We propose a classification framework for semantic type identification of compounds in Sanskrit. We broadly classify the compounds into four different classes namely, Avyayībhāva, Tatpuruṣa, Bahuvrīhi and Dvandva. Our classification is based on the traditional classification system followed by the ancient grammar treatise Adṣṭādhyāyī, proposed by Pāṇini 25 centuries back. We construct an elaborate features space for our system by combining conditional rules from the grammar Adṣṭādhyāyī, semantic relations between the compound components from a lexical database Amarakoṣa and linguistic structures from the data using Adaptor Grammars. Our in-depth analysis of the feature space highlight inadequacy of Adṣṭādhyāyī, a generative grammar, in classifying the data samples. Our experimental results validate the effectiveness of using lexical databases as suggested by Amba Kulkarni and Anil Kumar, and put forward a new research direction by introducing linguistic patterns obtained from Adaptor grammars for effective identification of compound type. We utilise an ensemble based approach, specifically designed for handling skewed datasets and we %and Experimenting with various classification methods, we achieve an overall accuracy of 0.77 using random forest classifiers.
Grapheme-to-Phoneme (G2P) conversion is the task of predicting the pronunciation of a word given its graphemic or written form. It is a highly important part of both automatic speech recognition (ASR) and text-to-speech (TTS) systems. In this paper, we evaluate seven G2P conversion approaches: Adaptive Regularization of Weight Vectors (AROW) based structured learning (S-AROW), Conditional Random Field (CRF), Joint-sequence models (JSM), phrase-based statistical machine translation (PBSMT), Recurrent Neural Network (RNN), Support Vector Machine (SVM) based point-wise classification, Weighted Finite-state Transducers (WFST) on a manually tagged Myanmar phoneme dictionary. The G2P bootstrapping experimental results were measured with both automatic phoneme error rate (PER) calculation and also manual checking in terms of voiced/unvoiced, tones, consonant and vowel errors. The result shows that CRF, PBSMT and WFST approaches are the best performing methods for G2P conversion on Myanmar language.
Named Entity Recognition (NER) is the task of classifying or labelling atomic elements in the text into categories such as Person, Location or Organisation. For Arabic language, recognizing named entities is a challenging task because of the complexity and the unique characteristics of this language. In addition, most of the previous work focuses on Modern Standard Arabic (MSA), however, recognizing named entities in social media is becoming more interesting these days. Dialectal Arabic (DA) and MSA are both used in social media, which is deemed as another challenging task. Most state-of-the-art Arabic NER systems count heavily on handcrafted engineering features and lexicons which is time consuming. In this paper, we introduce a novel neural network architecture which benefits both from character- and word-level representations automatically, by using combination of bidirectional LSTM and Conditional Random Field (CRF), eliminating the need for most feature engineering. Moreover, our model relies on unsupervised word representations learned from unannotated corpora. Experimental results demonstrate that our model achieves state-of-the-art performance on publicly available benchmark for Arabic NER for social media and surpassing the previous system by a large margin.
In recent years there has been a lot of interest in cross-lingual parsing for developing treebanks for languages with small or no annotated treebanks. In this paper, we explore the development of a cross-lingual transfer parser from Hindi to Bengali using a Hindi parser and a Hindi-Bengali parallel corpus. A parser is trained and applied to the Hindi sentences of the parallel corpus and the parse trees are projected to construct probable parse trees of the corresponding Bengali sentences. Only about 14% of these trees are complete (transferred trees contain all the target sentence words) and they are used to construct a Bengali parser. We relax the criteria of completeness to consider well-formed trees (43% of the trees) leading to an improvement. We note that the words often do not have a one-to-one mapping in the two languages but considering sentences at the chunk-level results in better correspondence between the two languages. Based on this we present a method to use chunking as a preprocessing step and do the transfer on the chunk trees. We find that about 72% of the projected parse trees of Bengali are now well-formed. The resultant parser achieves significant improvement in both Unlabeled Attachment Score (UAS) as well as Labeled Attachment Score (LAS) over the baseline word-level transferred parser.
Currently, corpus based-similarity, string-based similarity, and knowledge-based similarity techniques are used to compare short phrases. However, no work has been conducted on the similarity of phrases in Sinhala language. In this paper, we present a hybrid methodology to compute the similarity between two Sinhala sentences using a Semantic Similarity Measurement technique (corpus-based similarity measurement plus knowledge-based similarity measurement) that makes use of word order information. Since Sinhala WordNet is still under construction, we used lexical resources in performing this semantic similarity calculation. Evaluation using 4000 sentence pairs yielded an average MSE of 0.145 and a Pearson correla-tion factor of 0.832.
This paper focuses on the generation of case markers for free word order languages that use case markers as phrasal clitics for marking the relationship between the dependent-noun and its head. The generation of such clitics becomes essential task especially when translating from fixed word order languages where syntactic relations are identified by the positions of the dependent-nouns. To address the problem of missing markers on source-side, artificial markers are added in source to improve alignments with its target counterparts. Up to 1 BLEU point increase is observed over the baseline on different test sets for English-to-Urdu.
Action verbs are one of the frequently occurring linguistic elements in any given natural language as the speakers use them during every linguistic intercourse. However, each language expresses action verbs in its own inherently unique manner by categorization. One verb can refer to several interpretations of actions and one action can be expressed by more than one verb. The inter-language and intra-language variations create ambiguity for the translation of languages from the source language to target language with respect to action verbs. IMAGACT is a corpus-based ontological platform of action verbs translated from prototypic animated images explained in English and Italian as meta-languages. In this paper, we are presenting the issues and challenges in translating action verbs of Indian languages as target and English as source language by observing the animated images. Among the ten Indian languages which have been annotated so far on the platform are Sanskrit, Hindi, Urdu, Odia (Oriya), Bengali, Manipuri, Tamil, Assamese, Magahi and Marathi. Out of them, Manipuri belongs to the Sino-Tibetan, Tamil comes off the Dravidian and the rest owe their genesis to the Indo-Aryan language family. One of the issues is that the one-word morphological English verbs are translated into most of the Indian languages as verbs having more than one-word form; for instance as in the case of conjunct, compound, serial verbs and so on. We are further presenting a cross-lingual comparison of action verbs among Indian languages. In addition, we are also dealing with the issues in disambiguating animated images by the L1 native speakers using competence-based judgements and the theoretical and machine translation implications they bear.
The automatic analysis of emotions conveyed in social media content, e.g., tweets, has many beneficial applications. In the Philippines, one of the most disaster-prone countries in the world, such methods could potentially enable first responders to make timely decisions despite the risk of data deluge. However, recognising emotions expressed in Philippine-generated tweets, which are mostly written in Filipino, English or a mix of both, is a non-trivial task. In order to facilitate the development of natural language processing (NLP) methods that will automate such type of analysis, we have built a corpus of tweets whose predominant emotions have been manually annotated by means of crowdsourcing. Defining measures ensuring that only high-quality annotations were retained, we have produced a gold standard corpus of 1,146 emotion-labelled Filipino and English tweets. We validate the value of this manually produced resource by demonstrating that an automatic emotion-prediction method based on the use of a publicly available word-emotion association lexicon was unable to reproduce the labels assigned via crowdsourcing. While we are planning to make a few extensions to the corpus in the near future, its current version has been made publicly available in order to foster the development of emotion analysis methods based on advanced Filipino and English NLP.
In this paper, we describe the results of sentiment analysis on tweets in three Indian languages – Bengali, Hindi, and Tamil. We used the recently released SAIL dataset (Patra et al., 2015), and obtained state-of-the-art results in all three languages. Our features are simple, robust, scalable, and language-independent. Further, we show that these simple features provide better results than more complex and language-specific features, in two separate classification tasks. Detailed feature analysis and error analysis have been reported, along with learning curves for Hindi and Bengali.
In Machine Translation, divergence is one of the major barriers which plays a deciding role in determining the efficiency of the system at hand. Translation divergences originate when there is structural discrepancies between the input and the output languages. It can be of various types based on the issues we are addressing to such as linguistic, cultural, communicative and so on. Owing to the fact that two languages owe their origin to different language families, linguistic divergences emerge. The present study attempts at categorizing different types of linguistic divergences: the lexical-semantic and syntactic. In addition, it also helps identify and resolve the divergent linguistic features between English as source language and Bhojpuri as target language pair. Dorr’s theoretical framework (1994, 1994a) has been followed in the classification and resolution procedure. Furthermore, so far as the methodology is concerned, we have adhered to the Dorr’s Lexical Conceptual Structure for the resolution of divergences. This research will prove to be beneficial for developing efficient MT systems if the mentioned factors are incorporated considering the inherent structural constraints between source and target languages.ated considering the inherent structural constraints between SL and TL pairs.
In this paper, we discuss our creation of a web corpus of spoken Hindi (COSH), one of the Indo-Aryan languages spoken mainly in the Indian subcontinent. We also point out notable problems we’ve encountered in the web corpus and the special concordancer. After observing the kind of technical problems we encountered, especially regarding annotation tagged by Shiva Reddy’s tagger, we argue how they can be solved when using COSH for linguistic studies. Finally, we mention the kinds of linguistic research that we non-native speakers of Hindi can do using the corpus, especially in pragmatics and semantics, and from a comparative viewpoint to Japanese.
A sentence aligned parallel corpus is an important prerequisite in statistical machine translation. However, manual creation of such a parallel corpus is time consuming, and requires experts fluent in both languages. Automatic creation of a sentence aligned parallel corpus using parallel text is the solution to this problem. In this paper, we present the first ever empirical evaluation carried out to identify the best method to automatically create a sentence aligned Sinhala-Tamil parallel corpus. Annual reports from Sri Lankan government institutions were used as the parallel text for aligning. Despite both Sinhala and Tamil being under-resourced languages, we were able to achieve an F-score value of 0.791 using a hybrid approach that makes use of a bilingual dictionary.
Acquiring labeled speech for low-resource languages is a difficult task in the absence of native speakers of the language. One solution to this problem involves collecting speech transcriptions from crowd workers who are foreign or non-native speakers of a given target language. From these mismatched transcriptions, one can derive probabilistic phone transcriptions that are defined over the set of all target language phones using a noisy channel model. This paper extends prior work on deriving probabilistic transcriptions (PTs) from mismatched transcriptions by 1) modelling multilingual channels and 2) introducing a clustering-based phonetic mapping technique to improve the quality of PTs. Mismatched crowdsourcing for multilingual channels has certain properties of projection mapping, e.g., it can be interpreted as a clustering based on singular value decomposition of the segment alignments. To this end, we explore the use of distinctive feature weights, lexical tone confusions, and a two-step clustering algorithm to learn projections of phoneme segments from mismatched multilingual transcriber languages to the target language. We evaluate our techniques using mismatched transcriptions for Cantonese speech acquired from native English and Mandarin speakers. We observe a 5-9% relative reduction in phone error rate for the predicted Cantonese phone transcriptions using our proposed techniques compared with the previous PT method.
The paper describes a new tagset for the morphological disambiguation of Sanskrit, and compares the accuracy of two machine learning methods (Conditional Random Fields, deep recurrent neural networks) for this task, with a special focus on how to model the lexicographic information. It reports a significant improvement over previously published results.
In Cross-Language Information Retrieval, finding the appropriate translation of the source language query has always been a difficult problem to solve. We propose a technique towards solving this problem with the help of multilingual word clusters obtained from multilingual word embeddings. We use word embeddings of the languages projected to a common vector space on which a community-detection algorithm is applied to find clusters such that words that represent the same concept from different languages fall in the same group. We utilize these multilingual word clusters to perform query translation for Cross-Language Information Retrieval for three languages - English, Hindi and Bengali. We have experimented with the FIRE 2012 and Wikipedia datasets and have shown improvements over several standard methods like dictionary-based method, a transliteration-based model and Google Translate.
Neural machine translation (NMT) models have recently been shown to be very successful in machine translation (MT). The use of LSTMs in machine translation has significantly improved the translation performance for longer sentences by being able to capture the context and long range correlations of the sentences in their hidden layers. The attention model based NMT system (Bahdanau et al., 2014) has become the state-of-the-art, performing equal or better than other statistical MT approaches. In this paper, we wish to study the performance of the attention-model based NMT system (Bahdanau et al., 2014) on the Indian language pair, Hindi and Bengali, and do an analysis on the types or errors that occur in case when the languages are morphologically rich and there is a scarcity of large parallel training corpus. We then carry out certain post-processing heuristic steps to improve the quality of the translated statements and suggest further measures that can be carried out.
This paper presents a new comprehensive multi-level Part-Of-Speech tag set and a Support Vector Machine based Part-Of-Speech tagger for the Sinhala language. The currently available tag set for Sinhala has two limitations: the unavailability of tags to represent some word classes and the lack of tags to capture inflection based grammatical variations of words. The new tag set, presented in this paper overcomes both of these limitations. The accuracy of available Sinhala Part-Of-Speech taggers, which are based on Hidden Markov Models, still falls far behind state of the art. Our Support Vector Machine based tagger achieved an overall accuracy of 84.68% with 59.86% accuracy for unknown words and 87.12% for known words, when the test set contains 10% of unknown words.
Multilingual language processing tasks like statistical machine translation and cross language information retrieval rely mainly on availability of accurate parallel corpora. Manual construction of such corpus can be extremely expensive and time consuming. In this paper we present a simple yet efficient method to generate huge amount of reasonably accurate parallel corpus with minimal user efforts. We utilize the availability of large number of English books and their corresponding translations in other languages to build parallel corpus. Optical Character Recognizing systems are used to digitize such books. We propose a robust dictionary based parallel corpus generation system for alignment of multilingual text at different levels of granularity (sentence, paragraphs, etc). We show the performance of our proposed method on a manually aligned dataset of 300 Hindi-English sentences and 100 English-Malayalam sentences.
We present a research on learning Indonesian-Chinese bilingual lexicon using monolingual word embedding and bilingual seed lexicons to build shared bilingual word embedding space. We take the first attempt to examine the impact of different monolingual signals for the choice of seed lexicons on the model performance. We found that although monolingual signals alone do not seem to outperform signals coverings all words, the significant improvement for learning word translation of the same signal types may suggest that linguistic features possess value for further study in distinguishing the semantic margins of the shared word embedding space.
In this paper, we present how we generated two rich online bilingual dictionaries — Lao-French and French-Lao — from unstructured dictionaries in Microsoft Word files. Then we shortly discuss the possible reuse of the lexical data for Machine Translation projects.
The paper is a corpus study of the factors involved in disambiguating potential scope ambiguity in sentences with negation and universal quantifier, such as “I don’t want talk to all these people”, which can alternatively mean ‘I don’t want to talk to any of these people’ and ‘I don’t want to talk to some of these people’. The relevant factors are demonstrated to be largely different from those involved in disambiguating lexical polysemy. They include the syntactic function of the constituent containing “all” quantifier (subject, direct complement, adjunct), as well as the deepness of its embedding; the status of the main predicate and “all” constituent with respect to the information structure of the 6utterance (topic vs. focus, given vs. new information); pragmatic implicatures pertaining to the situations described in the utterances.
In this talk, I will outline a range of challenges presented by multiword expressions in terms of (lexicalist) precision grammar engineering, and different strategies for accommodating those challenges, in an attempt to strike the right balance in terms of generalisation and over- and under-generation.
Microsyntactic linguistic units, such as syntactic idioms and non-standard syntactic constructions, are poorly represented in linguistic resources, mostly because the former are elements occupying an intermediate position between the lexicon and the grammar and the latter are too specific to be routinely tackled by general grammars. Consequently, many such units produce substantial gaps in systems intended to solve sophisticated computational linguistics tasks, such as parsing, deep semantic analysis, question answering, machine translation, or text generation. They also present obstacles for applying advanced techniques to these tasks, such as machine learning. The paper discusses an approach aimed at bridging such gaps, focusing on the development of monolingual and multilingual corpora where microsyntactic units are to be tagged.
An excellent example of a phenomenon bridging a lexicon and a grammar is provided by grammaticalized alternations (e.g., passivization, reflexivity, and reciprocity): these alternations represent productive grammatical processes which are, however, lexically determined. While grammaticalized alternations keep lexical meaning of verbs unchanged, they are usually characterized by various changes in their morphosyntactic structure. In this contribution, we demonstrate on the example of reciprocity and its representation in the valency lexicon of Czech verbs, VALLEX how a linguistic description of complex (and still systemic) changes characteristic of grammaticalized alternations can benefit from an integration of grammatical rules into a valency lexicon. In contrast to other types of grammaticalized alternations, reciprocity in Czech has received relatively little attention although it closely interacts with various linguistic phenomena (e.g., with light verbs, diatheses, and reflexivity).
Language-endowed intelligent agents benefit from leveraging lexical knowledge falling at different points along a spectrum of compositionality. This means that robust computational lexicons should include not only the compositional expectations of argument-taking words, but also non-compositional collocations (idioms), semi-compositional collocations that might be difficult for an agent to interpret (e.g., standard metaphors), and even collocations that could be compositionally analyzed but are so frequently encountered that recording their meaning increases the efficiency of interpretation. In this paper we argue that yet another type of string-to-meaning mapping can also be useful to intelligent agents: remembered semantic analyses of actual text inputs. These can be viewed as super-specific multi-word expressions whose recorded interpretations mimic a person’s memories of knowledge previously learned from language input. These differ from typical annotated corpora in two ways. First, they provide a full, context-sensitive semantic interpretation rather than select features. Second, they are are formulated in the ontologically-grounded metalanguage used in a particular agent environment, meaning that the interpretations contribute to the dynamically evolving cognitive capabilites of agents configured in that environment.
Universal Dependencies is an initiative to develop cross-linguistically consistent grammatical annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning and parsing research from a language typology perspective. It assumes a dependency-based approach to syntax and a lexicalist approach to morphology, which together entail that the fundamental units of grammatical annotation are words. Words have properties captured by morphological annotation and enter into relations captured by syntactic annotation. Moreover, priority is given to relations between lexical content words, as opposed to grammatical function words. In this position paper, I discuss how this approach allows us to capture similarities and differences across typologically diverse languages.
Human communication is a multimodal activity, involving not only speech and written expressions, but intonation, images, gestures, visual clues, and the interpretation of actions through perception. In this paper, we describe the design of a multimodal lexicon that is able to accommodate the diverse modalities that present themselves in NLP applications. We have been developing a multimodal semantic representation, VoxML, that integrates the encoding of semantic, visual, gestural, and action-based features associated with linguistic expressions.
Valency slot filling is a semantic glue, which brings together the meanings of words. As regards the position of an argument in the dependency structure with respect to its predicate, there exist three types of valency filling: active (canonical), passive, and discontinuous. Of these, the first type is studied much better than the other two. As a rule, canonical actants are unambiguously marked in the syntactic structure, and each actant corresponds to a unique syntactic position. Linguistic information on which syntactic function an actant might have (subject, direct or indirect object), what its morphological form should be and which prepositions or conjunctions it requires, can be given in the lexicon in the form of government patterns, subcategorization frames, or similar data structures. We concentrate on non-canonical cases of valency filling in Russian, which are characteristic of non-verbal parts of speech, such as adverbs, adjectives, and particles, in the first place. They are more difficult to handle than canonical ones, because the position of the actant in the tree is governed by more complicated rules. A valency may be filled by expressions occupying different syntactic positions, and a syntactic position may accept expressions filling different valencies of the same word. We show how these phenomena can be processed in a semantic analyzer.
Verbenet is a French lexicon developed by “translation” of its English counterpart — VerbNet (Kipper-Schuler, 2005)—and treatment of the specificities of French syntax (Pradet et al., 2014; Danlos et al., 2016). One difficulty encountered in its development springs from the fact that the list of (potentially numerous) frames has no internal organization. This paper proposes a type system for frames that shows whether two frames are variants of a given alternation. Frame typing facilitates coherence checking of the resource in a “virtuous circle”. We present the principles underlying a program we developed and used to automatically type frames in VerbeNet. We also show that our system is portable to other languages.
We present an attempt to automatically identify Czech deverbative nouns using several methods that use large corpora as well as existing lexical resources. The motivation for the task is to extend a verbal valency (i.e., predicate-argument) lexicon by adding nouns that share the valency properties with the base verb, assuming their properties can be derived (even if not trivially) from the underlying verb by deterministic grammatical rules. At the same time, even in inflective languages, not all deverbatives are simply created from their underlying base verb by regular lexical derivation processes. We have thus developed hybrid techniques that use both large parallel corpora and several standard lexical resources. Thanks to the use of parallel corpora, the resulting sets contain also synonyms, which the lexical derivation rules cannot get. For evaluation, we have manually created a small, 100-verb gold data since no such dataset was initially available for Czech.
We present an interdisciplinary study on the interaction between the interpretation of noun-noun deverbal compounds (DCs; e.g., task assignment) and the morphosyntactic properties of their deverbal heads in English. Underlying hypotheses from theoretical linguistics are tested with tools and resources from computational linguistics. We start with Grimshaw’s (1990) insight that deverbal nouns are ambiguous between argument-supporting nominal (ASN) readings, which inherit verbal arguments (e.g., the assignment of the tasks), and the less verbal and more lexicalized Result Nominal and Simple Event readings (e.g., a two-page assignment). Following Grimshaw, our hypothesis is that the former will realize object arguments in DCs, while the latter will receive a wider range of interpretations like root compounds headed by non-derived nouns (e.g., chocolate box). Evidence from a large corpus assisted by machine learning techniques confirms this hypothesis, by showing that, besides other features, the realization of internal arguments by deverbal heads outside compounds (i.e., the most distinctive ASN-property in Grimshaw 1990) is a good predictor for an object interpretation of non-heads in DCs.
We show how to turn a large-scale syntactic dictionary into a dependency-based unification grammar where each piece of lexical information calls a separate rule, yielding a super granular grammar. Subcategorization, raising and control verbs, auxiliaries and copula, passivization, and tough-movement are discussed. We focus on the semantics-syntax interface and offer a new perspective on syntactic structure.
This paper presents our ongoing work on compilation of English multi-word expression (MWE) lexicon. We are especially interested in collecting flexible MWEs, in which some other components can intervene the expression such as “a number of” vs “a large number of” where a modifier of “number” can be placed in the expression and inherit the original meaning. We fiest collect possible candidates of flexible English MWEs from the web, and annotate all of their occurrences in the Wall Street Journal portion of Ontonotes corpus. We make use of word dependency strcuture information of the sentences converted from the phrase structure annotation. This process enables semi-automatic annotation of MWEs in the corpus and simultanaously produces the internal and external dependency representation of flexible MWEs.
The paper presents a contrastive description of reflexive possessive pronouns “svůj” in Czech and “svoj” in Russian. The research concerns syntactic, semantic and pragmatic aspects. With our analysis, we shed a new light on the already investigated issue, which comes from a detailed comparison of the phenomenon of possessive reflexivization in two typologically and genetically similar languages. We show that whereas in Czech, the possessive reflexivization is mostly limited to syntactic functions and does not go beyond the grammar, in Russian it gets additional semantic meanings and moves substan-tially towards the lexicon. The obtained knowledge allows us to explain heretofore unclear marginal uses of reflexives in each language.
A specific language as used by different speakers and in different situations has a number of more or less distant varieties. Extending the notion of non-standard language to varieties that do not fit an explicitly or implicitly assumed norm or pattern, we look for methods and tools that could be applied to this domain. The needs start from the theoretical side: categories usable for the analysis of non-standard language are not readily available, and continue to methods and tools required for its detection and diagnostics. A general discussion of issues related to non-standard language is followed by two case studies. The first study presents a taxonomy of morphosyntactic categories as an attempt to analyse non-standard forms produced by non-native learners of Czech. The second study focusses on the role of a rule-based grammar and lexicon in the process of building and using a parsebank.
Real world data differs radically from the benchmark corpora we use in NLP, resulting in large performance drops. The reason for this problem is obvious: NLP models are trained on limited samples from canonical varieties considered standard. However, there are many dimensions, e.g., sociodemographic, language, genre, sentence type, etc. on which texts can differ from the standard. The solution is not obvious: we cannot control for all factors, and it is not clear how to best go beyond the current practice of training on homogeneous data from a single domain and language. In this talk, I review the notion of canonicity, and how it shapes our community’s approach to language. I argue for the use of fortuitous data. Fortuitous data is data out there that just waits to be harvested. It includes data which is in plain sight, but is often neglected, and more distant sources like behavioral data, which first need to be refined. They provide additional contexts and a myriad of opportunities to build more adaptive language technology, some of which I will explore in this talk.
Entity linking and semantic parsing have been shown to be crucial to important applications such as question answering and document understanding. These tasks often require structured learning models, which make predictions on multiple interdependent variables. In this talk, I argue that carefully designed structured learning algorithms play a central role in entity linking and semantic parsing tasks. In particular, I will present several new structured learning models for entity linking, which jointly detect mentions and disambiguate entities as well as capture non-textual information. I will then show how to use a staged search procedure to building a state-of-the-art knowledge base question answering system. Finally, if time permits, I will discuss different supervision protocols for training semantic parsers and the value of labeling semantic parses.
This talk presents two NLP systems that were developed for helping disaster victims and rescue workers in the aftermath of large-scale disasters. DISAANA provides answers to questions such as “What is in short supply in Tokyo?” and displays locations related to each answer on a map. D-SUMM automatically summarizes a large number of disaster related reports concerning a specified area and helps rescue workers to understand disaster situations from a macro perspective. Both systems are publicly available as Web services. In the aftermath of the 2016 Kumamoto Earthquake (M7.0), the Japanese government actually used DISAANA to analyze the situation.
In this paper we present a series of experiments on discriminating between private and corporate accounts on Twitter. We define features based on Twitter metadata, morphosyntactic tags and surface forms, showing that the simple bag-of-words model achieves single best results that can, however, be improved by building a weighted soft ensemble of classifiers based on each feature type. Investigating the time and language dependence of each feature type delivers quite unexpecting results showing that features based on metadata are neither time- nor language-insensitive as the way the two user groups use the social network varies heavily through time and space.
User-generated content presents many challenges for its automatic processing. While many of them do come from out-of-vocabulary effects, others spawn from different linguistic phenomena such as unusual syntax. In this work we present a French three-domain data set made up of question headlines from a cooking forum, game chat logs and associated forums from two popular online games (MINECRAFT & LEAGUE OF LEGENDS). We chose these domains because they encompass different degrees of lexical and syntactic compliance with canonical language. We conduct an automatic and manual evaluation of the difficulties of processing these domains for part-of-speech prediction, and introduce a pilot study to determine whether dependency analysis lends itself well to annotate these data. We also discuss the development cost of our data set.
Information extraction from user-generated text has gained much attention with the growth of the Web.Disaster analysis using information from social media provides valuable, real-time, geolocation information for helping people caught up these in disasters. However, it is not convenient to analyze texts posted on social media because disaster keywords match any texts that contain words. For collecting posts about a disaster from social media, we need to develop a classifier to filter posts irrelevant to disasters. Moreover, because of the nature of social media, we can take advantage of posts that come with GPS information. However, a post does not always refer to an event occurring at the place where it has been posted. Therefore, we propose a new task of classifying whether a flood disaster occurred, in addition to predicting the geolocation of events from user-generated text. We report the annotation of the flood disaster corpus and develop a classifier to demonstrate the use of this corpus for disaster analysis.
We present a data-driven method for determining the veracity of a set of rumorous claims on social media data. Tweets from different sources pertaining to a rumor are processed on three levels: first, factuality values are assigned to each tweet based on four textual cue categories relevant for our journalism use case; these amalgamate speaker support in terms of polarity and commitment in terms of certainty and speculation. Next, the proportions of these lexical cues are utilized as predictors for tweet certainty in a generalized linear regression model. Subsequently, lexical cue proportions, predicted certainty, as well as their time course characteristics are used to compute veracity for each rumor in terms of the identity of the rumor-resolving tweet and its binary resolution value judgment. The system operates without access to extralinguistic resources. Evaluated on the data portion for which hand-labeled examples were available, it achieves .74 F1-score on identifying rumor resolving tweets and .76 F1-score on predicting if a rumor is resolved as true or false.
A major challenge for statistical machine translation (SMT) of Arabic-to-English user-generated text is the prevalence of text written in Arabizi, or Romanized Arabic. When facing such texts, a translation system trained on conventional Arabic-English data will suffer from extremely low model coverage. In addition, Arabizi is not regulated by any official standardization and therefore highly ambiguous, which prevents rule-based approaches from achieving good translation results. In this paper, we improve Arabizi-to-English machine translation by presenting a simple but effective Arabizi-to-Arabic transliteration pipeline that does not require knowledge by experts or native Arabic speakers. We incorporate this pipeline into a phrase-based SMT system, and show that translation quality after automatically transliterating Arabizi to Arabic yields results that are comparable to those achieved after human transliteration.
Name Variation in Community Question Answering Systems Abstract Community question answering systems are forums where users can ask and answer questions in various categories. Examples are Yahoo! Answers, Quora, and Stack Overflow. A common challenge with such systems is that a significant percentage of asked questions are left unanswered. In this paper, we propose an algorithm to reduce the number of unanswered questions in Yahoo! Answers by reusing the answer to the most similar past resolved question to the unanswered question, from the site. Semantically similar questions could be worded differently, thereby making it difficult to find questions that have shared needs. For example, “Who is the best player for the Reds?” and “Who is currently the biggest star at Manchester United?” have a shared need but are worded differently; also, “Reds” and “Manchester United” are used to refer to the soccer team Manchester United football club. In this research, we focus on question categories that contain a large number of named entities and entity name variations. We show that in these categories, entity linking can be used to identify relevant past resolved questions with shared needs as a given question by disambiguating named entities and matching these questions based on the disambiguated entities, identified entities, and knowledge base information related to these entities. We evaluated our algorithm on a new dataset constructed from Yahoo! Answers. The dataset contains annotated question pairs, (Qgiven, [Qpast, Answer]). We carried out experiments on several question categories and show that an entity-based approach gives good performance when searching for similar questions in entity rich categories.
Using aliases to refer to public figures is one way to make fun of people, to express sarcasm, or even to sidestep legal issues when expressing opinions on social media. However, linking an alias back to the real name is difficult, as it entails phonemic, graphemic, and semantic challenges. In this paper, we propose a phonemic-based approach and inject semantic information to align aliases with politicians’ Chinese formal names. The proposed approach creates an HMM model for each name to model its phonemes and takes into account document-level pairwise mutual information to capture the semantic relations to the alias. In this work we also introduce two new datasets consisting of 167 phonemic pairs and 279 mixed pairs of aliases and formal names. Experimental results show that the proposed approach models both phonemic and semantic information and outperforms previous work on both the phonemic and mixed datasets with the best top-1 accuracies of 0.78 and 0.59 respectively.
Accurate event detection in social media is very challenging because user generated contents are extremely noisy and sparse in content. Event indicators are generally words or phrases that act as a trigger that help us understand the semantics of the context they occur in. We present a weakly supervised approach that relies on using a single strong event indicator phrase as a seed to acquire a variety of additional event cues. We propose to leverage various types of implicit event indicators, such as props, actors and precursor events, to achieve precise event detection. We experimented with civil unrest events and show that the automatically learnt event indicators are effective in identifying specific types of events.
Stemming is an essential processing step in a wide range of high level text processing applications such as information extraction, machine translation and sentiment analysis. It is used to reduce words to their stems. Many stemming algorithms have been developed for Modern Standard Arabic (MSA). Although Arabic tweets and MSA are closely related and share many characteristics, there are substantial differences between them in lexicon and syntax. In this paper, we introduce a light Arabic stemmer for Arabic tweets. Our results show improvements over the performance of a number of well-known stemmers for Arabic.
Topic modelling techniques such as LDA have recently been applied to speech transcripts and OCR output. These corpora may contain noisy or erroneous texts which may undermine topic stability. Therefore, it is important to know how well a topic modelling algorithm will perform when applied to noisy data. In this paper we show that different types of textual noise can have diverse effects on the stability of topic models. On the other hand, topic model stability is not consistent with the same type but different levels of noise. We introduce a dictionary filtering approach to address this challenge, with the result that a topic model with the correct number of topics is always identified across different levels of noise.
Postmarketing surveillance (PMS) has the vital aim to monitor effects of drugs after release for use by the general population, but suffers from under-reporting and limited coverage. Automatic methods for detecting drug effect reports, especially for social media, could vastly increase the scope of PMS. Very few automatic PMS methods are currently available, in particular for the messy text types encountered on Twitter. In this paper we describe first results for developing PMS methods specifically for tweets. We describe the corpus of 125,669 tweets we have created and annotated to train and test the tools. We find that generic tools perform well for tweet-level language identification and tweet-level sentiment analysis (both 0.94 F1-Score). For detection of effect mentions we are able to achieve 0.87 F1-Score, while effect-level adverse-vs.-beneficial analysis proves harder with an F1-Score of 0.64. Among other things, our results indicate that MetaMap semantic types provide a very promising basis for identifying drug effect mentions in tweets.
In social networks services like Twitter, users are overwhelmed with huge amount of social data, most of which are short, unstructured and highly noisy. Identifying accurate information from this huge amount of data is indeed a hard task. Classification of tweets into organized form will help the user to easily access these required information. Our first contribution relates to filtering parts of speech and preprocessing this kind of highly noisy and short data. Our second contribution concerns the named entity recognition (NER) in tweets. Thus, the adaptation of existing language tools for natural languages, noisy and not accurate language tweets, is necessary. Our third contribution involves segmentation of hashtags and a semantic enrichment using a combination of relations from WordNet, which helps the performance of our classification system, including disambiguation of named entities, abbreviations and acronyms. Graph theory is used to cluster the words extracted from WordNet and tweets, based on the idea of connected components. We test our automatic classification system with four categories: politics, economy, sports and the medical field. We evaluate and compare several automatic classification systems using part or all of the items described in our contributions and found that filtering by part of speech and named entity recognition dramatically increase the classification precision to 77.3 %. Moreover, a classification system incorporating segmentation of hashtags and semantic enrichment by two relations from WordNet, synonymy and hyperonymy, increase classification precision up to 83.4 %.
Text normalization techniques based on rules, lexicons or supervised training requiring large corpora are not scalable nor domain interchangeable, and this makes them unsuitable for normalizing user-generated content (UGC). Current tools available for Brazilian Portuguese make use of such techniques. In this work we propose a technique based on distributed representation of words (or word embeddings). It generates continuous numeric vectors of high-dimensionality to represent words. The vectors explicitly encode many linguistic regularities and patterns, as well as syntactic and semantic word relationships. Words that share semantic similarity are represented by similar vectors. Based on these features, we present a totally unsupervised, expandable and language and domain independent method for learning normalization lexicons from word embeddings. Our approach obtains high correction rate of orthographic errors and internet slang in product reviews, outperforming the current available tools for Brazilian Portuguese.
The SemEval-2010 benchmark dataset has brought renewed attention to the task of automatic keyphrase extraction. This dataset is made up of scientific articles that were automatically converted from PDF format to plain text and thus require careful preprocessing so that irrevelant spans of text do not negatively affect keyphrase extraction performance. In previous work, a wide range of document preprocessing techniques were described but their impact on the overall performance of keyphrase extraction models is still unexplored. Here, we re-assess the performance of several keyphrase extraction models and measure their robustness against increasingly sophisticated levels of document preprocessing.
Text normalization is the task of transforming lexical variants to their canonical forms. We model the problem of text normalization as a character-level sequence to sequence learning problem and present a neural encoder-decoder model for solving it. To train the encoder-decoder model, many sentences pairs are generally required. However, Japanese non-standard canonical pairs are scarce in the form of parallel corpora. To address this issue, we propose a method of data augmentation to increase data size by converting existing resources into synthesized non-standard forms using handcrafted rules. We conducted an experiment to demonstrate that the synthesized corpus contributes to stably train an encoder-decoder model and improve the performance of Japanese text normalization.
This paper presents the results of the Twitter Named Entity Recognition shared task associated with W-NUT 2016: a named entity tagging task with 10 teams participating. We outline the shared task, annotation process and dataset statistics, and provide a high-level overview of the participating systems for each shared task.
In this paper, we present our approach for named entity recognition in Twitter messages that we used in our participation in the Named Entity Recognition in Twitter shared task at the COLING 2016 Workshop on Noisy User-generated text (WNUT). The main challenge that we aim to tackle in our participation is the short, noisy and colloquial nature of tweets, which makes named entity recognition in Twitter message a challenging task. In particular, we investigate an approach for dealing with this problem by enabling bidirectional long short-term memory (LSTM) to automatically learn orthographic features without requiring feature engineering. In comparison with other systems participating in the shared task, our system achieved the most effective performance on both the ‘segmentation and categorisation’ and the ‘segmentation only’ sub-tasks.
Named entity recognition (NER) in social media (e.g., Twitter) is a challenging task due to the noisy nature of text. As part of our participation in the W-NUT 2016 Named Entity Recognition Shared Task, we proposed an unsupervised learning approach using deep neural networks and leverage a knowledge base (i.e., DBpedia) to bootstrap sparse entity types with weakly labelled data. To further boost the performance, we employed a more sophisticated tagging scheme and applied dropout as a regularisation technique in order to reduce overfitting. Even without hand-crafting linguistic features nor leveraging any of the W-NUT-provided gazetteers, we obtained robust performance with our approach, which ranked third amongst all shared task participants according to the official evaluation on a gold standard named entity-annotated corpus of 3,856 tweets.
Twitter named entity recognition is the process of identifying proper names and classifying them into some predefined labels/categories. The paper introduces a Twitter named entity system using a supervised machine learning approach, namely Conditional Random Fields. A large set of different features was developed and the system was trained using these. The Twitter named entity task can be divided into two parts: i) Named entity extraction from tweets and ii) Twitter name classification into ten different types. For Twitter named entity recognition on unseen test data, our system obtained the second highest F1 score in the shared task: 63.22%. The system performance on the classification task was worse, with an F1 measure of 40.06% on unseen test data, which was the fourth best of the ten systems participating in the shared task.
We presented in this work our participation in the 2nd Named Entity Recognition for Twitter shared task. The task has been cast as a sequence labeling one and we employed a learning to search approach in order to tackle it. We also leveraged LOD for extracting rich contextual features for the named-entities. Our submission achieved F-scores of 46.16 and 60.24 for the classification and the segmentation tasks and ranked 2nd and 3rd respectively. The post-analysis showed that LOD features improved substantially the performance of our system as they counter-balance the lack of context in tweets. The shared task gave us the opportunity to test the performance of NER systems in short and noisy textual data. The results of the participated systems shows that the task is far to be considered as a solved one and methods with stellar performance in normal texts need to be revised.
In this paper, we describe the DeepNNNER entry to The 2nd Workshop on Noisy User-generated Text (WNUT) Shared Task #2: Named Entity Recognition in Twitter. Our shared task submission adopts the bidirectional LSTM-CNN model of Chiu and Nichols (2016), as it has been shown to perform well on both newswire and Web texts. It uses word embeddings trained on large-scale Web text collections together with text normalization to cope with the diversity in Web texts, and lexicons for target named entity classes constructed from publicly-available sources. Extended evaluation comparing the effectiveness of various word embeddings, text normalization, and lexicon settings shows that our system achieves a maximum F1-score of 47.24, performance surpassing that of the shared task’s second-ranked system.
This paper describes the ASU system submitted in the COLING W-NUT 2016 Twitter Named Entity Recognition (NER) task. We present an experimental study on applying deep learning to extracting named entities (NEs) from tweets. We built two Long Short-Term Memory (LSTM) models for the task. The first model was built to extract named entities without types while the second model was built to extract and then classify them into 10 fine-grained entity classes. In this effort, we show detailed experimentation results on the effectiveness of word embeddings, brown clusters, part-of-speech (POS) tags, shape features, gazetteers, and local context for the tweet input vector representation to the LSTM model. Also, we present a set of experiments, to better design the network parameters for the Twitter NER task. Our system was ranked the fifth out of ten participants with a final f1-score for the typed classes of 39% and 55% for the non typed ones.
This paper describes our system used in the 2nd Workshop on Noisy User-generated Text (WNUT) shared task for Named Entity Recognition (NER) in Twitter, in conjunction with Coling 2016. Our system is based on supervised machine learning by applying Conditional Random Fields (CRF) to train two classifiers for two evaluations. The first evaluation aims at predicting the 10 fine-grained types of named entities; while the second evaluation aims at predicting no type of named entities. The experimental results show that our method has significantly improved Twitter NER performance.
Many of the existing Named Entity Recognition (NER) solutions are built based on news corpus data with proper syntax. These solutions might not lead to highly accurate results when being applied to noisy, user generated data, e.g., tweets, which can feature sloppy spelling, concept drift, and limited contextualization of terms and concepts due to length constraints. The models described in this paper are based on linear chain conditional random fields (CRFs), use the BIEOU encoding scheme, and leverage random feature dropout for up-sampling the training data. The considered features include word clusters and pre-trained distributed word representations, updated gazetteer features, and global context predictions. The latter feature allows for ingesting the meaning of new or rare tokens into the system via unsupervised learning and for alleviating the need to learn lexicon based features, which usually tend to be high dimensional. In this paper, we report on the solution [ST] we submitted to the WNUT 2016 NER shared task. We also present an improvement over our original submission [SI], which we built by using semi-supervised learning on labelled training data and pre-trained resourced constructed from unlabelled tweet data. Our ST solution achieved an F1 score of 1.2% higher than the baseline (35.1% F1) for the task of extracting 10 entity types. The SI resulted in an increase of 8.2% in F1 score over the base-line (7.08% over ST). Finally, the SI model’s evaluation on the test data achieved a F1 score of 47.3% (~1.15% increase over the 2nd best submitted solution). Our experimental setup and results are available as a standalone twitter NER tool at https://github.com/napsternxg/TwitterNER.
This paper presents the shared task for English Twitter geolocation prediction in WNUT 2016. We discuss details of task settings, data preparations and participant systems. The derived dataset and performance figures from each system provide baselines for future research in this realm.
In this paper, we describe CSIRO Data61’s participation in the Geolocation shared task at the Workshop for Noisy User-generated Text. Our approach was to use ensemble methods to capitalise on four component methods: heuristics based on metadata, a label propagation method, timezone text classifiers, and an information retrieval approach. The ensembles we explored focused on examining the role of language technologies in geolocation prediction and also in examining the use of hard voting and cascading ensemble methods. Based on the accuracy of city-level predictions, our systems were the best performing submissions at this year’s shared task. Furthermore, when estimating the latitude and longitude of a user, our median error distance was accurate to within 30 kilometers.
Knowing the location of a social media user and their posts is important for various purposes, such as the recommendation of location-based items/services, and locality detection of crisis/disasters. This paper describes our submission to the shared task “Geolocation Prediction in Twitter” of the 2nd Workshop on Noisy User-generated Text. In this shared task, we propose an algorithm to predict the location of Twitter users and tweets using a multinomial Naive Bayes classifier trained on Location Indicative Words and various textual features (such as city/country names, #hashtags and @mentions). We compared our approach against various baselines based on Location Indicative Words, city/country names, #hashtags and @mentions as individual feature sets, and experimental results show that our approach outperforms these baselines in terms of classification accuracy, mean and median error distance.
This paper describes a model that we submitted to W-NUT 2016 Shared task #1: Geolocation Prediction in Twitter. Our model classifies a tweet or a user to a city using a simple neural networks structure with fully-connected layers and average pooling processes. From the findings of previous geolocation prediction approaches, we integrated various user metadata along with message texts and trained the model with them. In the test run of the task, the model achieved the accuracy of 40.91% and the median distance error of 69.50 km in message-level prediction and the accuracy of 47.55% and the median distance error of 16.13 km in user-level prediction. These results are moderate performances in terms of accuracy and best performances in terms of distance. The results show a promising extension of neural networks based models for geolocation prediction where recent advances in neural networks can be added to enhance our current simple model.
The availability of Language Technology Resources and Tools generates a considerable methodological potential in the Digital Humanities: aspects of research questions from the Humanities and Social Sciences can be addressed on text collections in ways that were unavailable to traditional approaches. I start this talk by sketching some sample scenarios of Digital Humanities projects which involve various Humanities and Social Science disciplines, noting that the potential for a meaningful contribution to higher-level questions is highest when the employed language technological models are carefully tailored both (a) to characteristics of the given target corpus, and (b) to relevant analytical subtasks feeding the discipline-specific research questions. Keeping up a multidisciplinary perspective, I then point out a recurrent dilemma in Digital Humanities projects that follow the conventional set-up of collaboration: to build high-quality computational models for the data, fixed analytical targets should be specified as early as possible – but to be able to respond to Humanities questions as they evolve over the course of analysis, the analytical machinery should be kept maximally flexible. To reach both, I argue for a novel collaborative culture that rests on a more interleaved, continuous dialogue. (Re-)Specification of analytical targets should be an ongoing process in which the Humanities Scholars and Social Scientists play a role that is as important as the Computational Scientists’ role. A promising approach lies in the identification of re-occurring types of analytical subtasks, beyond linguistic standard tasks, which can form building blocks for text analysis across disciplines, and for which corpus-based characterizations (viz. annotations) can be collected, compared and revised. On such grounds, computational modeling is more directly tied to the evolving research questions, and hence the seemingly opposing needs of reliable target specifications vs. “malleable” frameworks of analysis can be reconciled. Experimental work following this approach is under way in the Center for Reflected Text Analytics (CRETA) in Stuttgart.
We examine two different methods for finding rising words (among which neologisms) and falling words (among which archaisms) in decades of magazine texts (millions of words) and in years of tweets (billions of words): one based on correlation coefficients of relative frequencies and time, and one based on comparing initial and final word frequencies of time intervals. We find that smoothing frequency scores improves the precision scores of both methods and that the correlation coefficients perform better on magazine text but worse on tweets. Since the two ranking methods find different words they can be used in side-by-side to study the behavior of words over time.
Multimodal question answering in the cultural heritage domain allows visitors to ask questions in a more natural way and thus provides better user experiences with cultural objects while visiting a museum, landmark or any other historical site. In this paper, we introduce the construction of a golden standard dataset that will aid research of multimodal question answering in the cultural heritage domain. The dataset, which will be soon released to the public, contains multimodal content including images of typical artworks from the fascinating old-Egyptian Amarna period, related image-containing documents of the artworks and over 800 multimodal queries integrating visual and textual questions. The multimodal questions and related documents are all in English. The multimodal questions are linked to relevant paragraphs in the related documents that contain the answer to the multimodal query.
In this paper a social network is extracted from a literary text. The social network shows, how frequent the characters interact and how similar their social behavior is. Two types of similarity measures are used: the first applies co-occurrence statistics, while the second exploits cosine similarity on different types of word embedding vectors. The results are evaluated by a paid micro-task crowdsourcing survey. The experiments suggest that specific types of word embeddings like word2vec are well-suited for the task at hand and the specific circumstances of literary fiction text.
We present an approach to detect differences in lexical semantics across English language registers, using word embedding models from distributional semantics paradigm. Models trained on register-specific subcorpora of the BNC corpus are employed to compare lists of nearest associates for particular words and draw conclusions about their semantic shifts depending on register in which they are used. The models are evaluated on the task of register classification with the help of the deep inverse regression approach. Additionally, we present a demo web service featuring most of the described models and allowing to explore word meanings in different English registers and to detect register affiliation for arbitrary texts. The code for the service can be easily adapted to any set of underlying models.
We are constructing an annotated diachronic corpora of the Japanese language. In part of thiswork, we construct a corpus of Manyosyu, which is an old Japanese poetry anthology. In thispaper, we describe how to align the transcribed text and its original text semiautomatically to beable to cross-reference them in our Manyosyu corpus. Although we align the original charactersto the transcribed words manually, we preliminarily align the transcribed and original charactersby using an unsupervised automatic alignment technique of statistical machine translation toalleviate the work. We found that automatic alignment achieves an F1-measure of 0.83; thus, each poem has 1–2 alignment errors. However, finding these errors and modifying them are less workintensiveand more efficient than fully manual annotation. The alignment probabilities can beutilized in this modification. Moreover, we found that we can locate the uncertain transcriptionsin our corpus and compare them to other transcriptions, by using the alignment probabilities.
Arabic is a widely-spoken language with a rich and long history spanning more than fourteen centuries. Yet existing Arabic corpora largely focus on the modern period or lack sufficient diachronic information. We develop a large-scale, historical corpus of Arabic of about 1 billion words from diverse periods of time. We clean this corpus, process it with a morphological analyzer, and enhance it by detecting parallel passages and automatically dating undated texts. We demonstrate its utility with selected case-studies in which we show its application to the digital humanities.
We here describe a novel methodology for measuring affective language in historical text by expanding an affective lexicon and jointly adapting it to prior language stages. We automatically construct a lexicon for word-emotion association of 18th and 19th century German which is then validated against expert ratings. Subsequently, this resource is used to identify distinct emotional patterns and trace long-term emotional trends in different genres of writing spanning several centuries.
Historical treebanks tend to be manually annotated, which is not surprising, since state-of-the-art parsers are not accurate enough to ensure high-quality annotation for historical texts. We test whether automatic parsing can be an efficient pre-annotation tool for Old East Slavic texts. We use the TOROT treebank from the PROIEL treebank family. We convert the PROIEL format to the CONLL format and use MaltParser to create syntactic pre-annotation. Using the most conservative evaluation method, which takes into account PROIEL-specific features, MaltParser by itself yields 0.845 unlabelled attachment score, 0.779 labelled attachment score and 0.741 secondary dependency accuracy (note, though, that the test set comes from a relatively simple genre and contains rather short sentences). Experiments with human annotators show that preparsing, if limited to sentences where no changes to word or sentence boundaries are required, increases their annotation rate. For experienced annotators, the speed gain varies from 5.80% to 16.57%, for inexperienced annotators from 14.61% to 32.17% (using conservative estimates). There are no strong reliable differences in the annotation accuracy, which means that there is no reason to suspect that using preparsing might lower the final annotation quality.
In this paper we will discuss a method for data visualization together with its potential usefulness in digital humanities and philosophy of language. We compiled a multilingual parallel corpus from different versions of Wittgenstein’s Tractatus Logico-philosophicus, including the original in German and translations into English, Spanish, French, and Russian. Using this corpus, we compute a similarity measure between propositions and render a visual network of relations for different languages.
We introduce the third major release of WebAnno, a generic web-based annotation tool for distributed teams. New features in this release focus on semantic annotation tasks (e.g. semantic role labelling or event annotation) and allow the tight integration of semantic annotations with syntactic annotations. In particular, we introduce the concept of slot features, a novel constraint mechanism that allows modelling the interaction between semantic and syntactic annotations, as well as a new annotation user interface. The new features were developed and used in an annotation project for semantic roles on German texts. The paper briefly introduces this project and reports on experiences performing annotations with the new tool. On a comparative evaluation, our tool reaches significant speedups over WebAnno 2 for a semantic annotation task.
Although spanning thousands of years and genres as diverse as liturgy, historiography, lyric and other forms of prose and poetry, the body of Latin texts is still relatively sparse compared to English. Data sparsity in Latin presents a number of challenges for traditional Named Entity Recognition techniques. Solving such challenges and enabling reliable Named Entity Recognition in Latin texts can facilitate many down-stream applications, from machine translation to digital historiography, enabling Classicists, historians, and archaeologists for instance, to track the relationships of historical persons, places, and groups on a large scale. This paper presents the first annotated corpus for evaluating Named Entity Recognition in Latin, as well as a fully supervised model that achieves over 90% F-score on a held-out test set, significantly outperforming a competitive baseline. We also present a novel active learning strategy that predicts how many and which sentences need to be annotated for named entities in order to attain a specified degree of accuracy when recognizing named entities automatically in a given text. This maximizes the productivity of annotators while simultaneously controlling quality.
We present ANNISVis, a webapp for comparative visualization of geographical distribution of linguistic data, as well as a sample deployment for a corpus of Middle High German texts. Unlike existing geographical visualization solutions, which work with pre-existing data sets, or are bound to specific corpora, ANNISVis allows the user to formulate multiple ad-hoc queries and visualizes them on a map, and it can be configured for any corpus that can be imported into ANNIS. This enables explorative queries of the quantitative aspects of a corpus with geographical features. The tool will be made available to download in open source.
In the Danish CLARIN-DK infrastructure, chaining language technology (LT) tools into a workflow is easy even for a non-expert user, because she only needs to specify the input and the desired output of the workflow. With this information and the registered input and output profiles of the available tools, the CLARIN-DK workflow management system (WMS) computes combinations of tools that will give the desired result. This advanced functionality was originally not envisaged, but came within reach by writing the WMS partly in Java and partly in a programming language for symbolic computation, Bracmat. Handling LT tool profiles, including the computation of workflows, is easier with Bracmat’s language constructs for tree pattern matching and tree construction than with the language constructs offered by mainstream programming languages.
Machine Translation (MT) plays a critical role in expanding capacity in the translation industry. However, many valuable documents, including digital documents, are encoded in non-accessible formats for machine processing (e.g., Historical or Legal documents). Such documents must be passed through a process of Optical Character Recognition (OCR) to render the text suitable for MT. No matter how good the OCR is, this process introduces recognition errors, which often renders MT ineffective. In this paper, we propose a new OCR to MT framework based on adding a new OCR error correction module to enhance the overall quality of translation. Experimentation shows that our new system correction based on the combination of Language Modeling and Translation methods outperforms the baseline system by nearly 30% relative improvement.
In this paper we describe how the complexity of human communication can be analysed with the help of language technology. We present the HuComTech corpus, a multimodal corpus containing 50 hours of videotaped interviews containing a rich annotation of about 2 million items annotated on 33 levels. The corpus serves as a general resource for a wide range of re-search addressing natural conversation between humans in their full complexity. It can benefit particularly digital humanities researchers working in the field of pragmatics, conversational analysis and discourse analysis. We will present a number of tools and automated methods that can help such enquiries. In particular, we will highlight the tool Theme, which is designed to uncover hidden temporal patterns (called T-patterns) in human interaction, and will show how it can applied to the study of multimodal communication.
Most modern and post-modern poems have developed a post-metrical idea of lyrical prosody that employs rhythmical features of everyday language and prose instead of a strict adherence to rhyme and metrical schemes. This development is subsumed under the term free verse prosody. We present our methodology for the large-scale analysis of modern and post-modern poetry in both their written form and as spoken aloud by the author. We employ language processing tools to align text and speech, to generate a null-model of how the poem would be spoken by a naïve reader, and to extract contrastive prosodic features used by the poet. On these, we intend to build our model of free verse prosody, which will help to understand, differentiate and relate the different styles of free verse poetry. We plan to use our processing scheme on large amounts of data to iteratively build models of styles, to validate and guide manual style annotation, to identify further rhythmical categories, and ultimately to broaden our understanding of free verse poetry. In this paper, we report on a proof-of-concept of our methodology using smaller amounts of poems and a limited set of features. We find that our methodology helps to extract differentiating features in the authors’ speech that can be explained by philological insight. Thus, our automatic method helps to guide the literary analysis and this in turn helps to improve our computational models.
This paper presents a tool to investigate the design of multimodal instructions (MIs), i.e., instructions that contain both text and pictures. The benefit of including pictures in information presentation has been established, but the characteristics of those pictures and of their textual counterparts and the rela-tion(s) between them have not been researched in a systematic manner. We present the PAT Work-bench, a tool to store, annotate and retrieve MIs based on a validated coding scheme with currently 42 categories that describe instructions in terms of textual features, pictorial elements, and relations be-tween text and pictures. We describe how the PAT Workbench facilitates collaborative annotation and inter-annotator agreement calculation. Future work on the tool includes expanding its functionality and usability by (i) making the MI annotation scheme dynamic for adding relevant features based on empirical evaluations of the MIs, (ii) implementing algorithms for automatic tagging of MI features, and (iii) implementing automatic MI evaluation algorithms based on results obtained via e.g. crowdsourced assessments of MIs.
The increasing amount of multilingual text collections available in different domains makes its automatic processing essential for the development of a given field. However, standard processing techniques based on statistical clues and keyword searches have clear limitations. Instead, we propose a knowledge-based processing pipeline which overcomes most of the limitations of these techniques. This, in turn, enables direct comparison across texts in different languages without the need of translation. In this paper we show the potential of this approach for semantically indexing multilingual text collections in the history domain. In our experiments we used a version of the Bible translated in four different languages, evaluating the precision of our semantic indexing pipeline and showing its reliability on the cross-lingual text retrieval task.
This paper presents on-going work on creating NLP tools for under-resourced languages from very sparse training data coming from linguistic field work. In this work, we focus on Ingush, a Nakh-Daghestanian language spoken by about 300,000 people in the Russian republics Ingushetia and Chechnya. We present work on morphosyntactic taggers trained on transcribed and linguistically analyzed recordings and dependency parsers using English glosses to project annotation for creating synthetic treebanks. Our preliminary results are promising, supporting the goal of bootstrapping efficient NLP tools with limited or no task-specific annotated data resources available.
This paper gives an overview of the MultiTal project, which aims to create a research infrastructure that ensures long-term distribution of NLP tools descriptions. The goal is to make NLP tools more accessible and usable to end-users of different disciplines. The infrastructure is built on a meta-data scheme modelling and standardising multilingual NLP tools documentation. The model is conceptualised using an OWL ontology. The formal representation of the ontology allows us to automatically generate organised and structured documentation in different languages for each represented tool.
This article describes work on enabling the addition of temporal information to senses of words in linguistic linked open data lexica based on the lemonDia model. Our contribution in this article is twofold. On the one hand, we demonstrate how lemonDia enables the querying of diachronic lexical datasets using OWL-oriented Semantic Web based technologies. On the other hand, we present a preliminary version of an interactive interface intended to help users in creating lexical datasets that model meaning change over time.
(This is the abstract for the submission.) Large-scale comparisons between the poetry of Tang and Song dynasties shed light on how words and expressions were used and shared among the poets. That some words were used only in the Tang poetry and some only in the Song poetry could lead to interesting research in linguistics. That the most frequent colors are different in the Tang and Song poetry provides a trace of the changing social circumstances in the dynasties. Results of the current work link to research topics of lexicography, semantics, and social transitions. We discuss our findings and present our algorithms for efficient comparisons among the poems, which are crucial for completing billion times of comparisons within acceptable time.
The following paper describes the first steps in the development of an ontology for the textbook research discipline. The aim of the project WorldViews is to establish a digital edition focussing on views of the world depicted in textbooks. For this purpose an initial TEI profile has been formalised and tested as a use case to enable the semantical encoding of the resource ‘textbook’. This profile shall provide a basic data model describing major facets of the textbook’s structure relevant to historians.
In this paper we present a new combination of existing language tools for Polish with a popular data mining platform intended to help researchers from digital humanities perform computational analyses without any programming. The toolset includes RapidMiner Studio, a software solution offering graphical setup of integrated analytical processes and Multiservice, a Web service offering access to several state-of-the-art linguistic tools for Polish. The setting is verified in a simple task of counting frequencies of unknown words in a small corpus.
In this paper, we propose to use a subfield of machine learning –grammatical inference– to measure linguistic complexity from a developmental point of view. We focus on relative complexity by considering a child learner in the process of first language acquisition. The relevance of grammatical inference models for measuring linguistic complexity from a developmental point of view is based on the fact that algorithms proposed in this area can be considered computational models for studying first language acquisition. Even though it will be possible to use different techniques from the field of machine learning as computational models for dealing with linguistic complexity -since in any model we have algorithms that can learn from data-, we claim that grammatical inference models offer some advantages over other tools.
In this paper, we introduce for the first time a Distributional Model for computing semantic complexity, inspired by the general principles of the Memory, Unification and Control framework(Hagoort, 2013; Hagoort, 2016). We argue that sentence comprehension is an incremental process driven by the goal of constructing a coherent representation of the event represented by the sentence. The composition cost of a sentence depends on the semantic coherence of the event being constructed and on the activation degree of the linguistic constructions. We also report the results of a first evaluation of the model on the Bicknell dataset (Bicknell et al., 2010).
We present a novel approach to the automatic assessment of text complexity based on a sliding-window technique that tracks the distribution of complexity within a text. Such distribution is captured by what we term “complexity contours” derived from a series of measurements for a given linguistic complexity measure. This approach is implemented in an automatic computational tool, CoCoGen – Complexity Contour Generator, which in its current version supports 32 indices of linguistic complexity. The goal of the paper is twofold: (1) to introduce the design of our computational tool based on a sliding-window technique and (2) to showcase this approach in the area of second language (L2) learning, i.e. more specifically, in the area of L2 writing.
This study demonstrates a weakness in how n-gram and PCFG surprisal are used to predict reading times in eye-tracking data. In particular, the information conveyed by words skipped during saccades is not usually included in the surprisal measures. This study shows that correcting the surprisal calculation improves n-gram surprisal and that upcoming n-grams affect reading times, replicating previous findings of how lexical frequencies affect reading times. In contrast, the predictivity of PCFG surprisal does not benefit from the surprisal correction despite the fact that lexical sequences skipped by saccades are processed by readers, as demonstrated by the corrected n-gram measure. These results raise questions about the formulation of information-theoretic measures of syntactic processing such as PCFG surprisal and entropy reduction when applied to reading times.
Computational approaches to readability assessment are generally built and evaluated using gold standard corpora labeled by publishers or teachers rather than being grounded in observations about human performance. Considering that both the reading process and the outcome can be observed, there is an empirical wealth that could be used to ground computational analysis of text readability. This will also support explicit readability models connecting text complexity and the reader’s language proficiency to the reading process and outcomes. This paper takes a step in this direction by reporting on an experiment to study how the relation between text complexity and reader’s language proficiency affects the reading process and performance outcomes of readers after reading We modeled the reading process using three eye tracking variables: fixation count, average fixation count, and second pass reading duration. Our models for these variables explained 78.9%, 74% and 67.4% variance, respectively. Performance outcome was modeled through recall and comprehension questions, and these models explained 58.9% and 27.6% of the variance, respectively. While the online models give us a better understanding of the cognitive correlates of reading with text complexity and language proficiency, modeling of the offline measures can be particularly relevant for incorporating user aspects into readability models.
Studies on the role of memory as a predictor of reading time latencies (1) differ in their predictions about when memory effects should occur in processing and (2) have had mixed results, with strong positive effects emerging from isolated constructed stimuli and weak or even negative effects emerging from naturally-occurring stimuli. Our study addresses these concerns by comparing several implementations of prominent sentence processing theories on an exploratory corpus and evaluating the most successful of these on a confirmatory corpus, using a new self-paced reading corpus of seemingly natural narratives constructed to contain an unusually high proportion of memory-intensive constructions. We show highly significant and complementary broad-coverage latency effects both for predictors based on the Dependency Locality Theory and for predictors based on a left-corner parsing model of sentence processing. Our results indicate that memory access during sentence processing does take time, but suggest that stimuli requiring many memory access events may be necessary in order to observe the effect.
Lexical complexity plays a central role in readability, particularly for dyslexic children and poor readers because of their slow and laborious decoding and word recognition skills. Although some features to aid readability may be common to most languages (e.g., the majority of ‘easy’ words are of low frequency), we believe that lexical complexity is mainly language-specific. In this paper, we define lexical complexity for French and we present a pilot study on the effects of text simplification in dyslexic children. The participants were asked to read out loud original and manually simplified versions of a standardized French text corpus and to answer comprehension questions after reading each text. The analysis of the results shows that the simplifications performed were beneficial in terms of reading speed and they reduced the number of reading errors (mainly lexical ones) without a loss in comprehension. Although the number of participants in this study was rather small (N=10), the results are promising and contribute to the development of applications in computational linguistics.
In this paper we will be dealing with different levels of complexity in the processing of Italian, a Romance language inheriting many properties from Latin which make it an almost free word order language . The paper is concerned with syntactic complexity as measurable on the basis of the cognitive parser that incrementally builds up a syntactic representation to be used by the semantic component. The theory behind will be LFG and parsing preferences will be used to justify one choice both from a principled and a processing point of view. LFG is a transformationless theory in which there is no deep structure separate from surface syntactic structure. This is partially in accordance with constructional theories in which noncanonical structures containing non-argument functions FOCUS/TOPIC are treated as multifunctional constituents. Complexity is computed on a processing basis following suggestions made by Blache and demonstrated by Kluender and Chesi
Previous researches have shown that learning multiple representations for polysemous words can improve the performance of word embeddings on many tasks. However, this leads to another problem. Several vectors of a word may actually point to the same meaning, namely pseudo multi-sense. In this paper, we introduce the concept of pseudo multi-sense, and then propose an algorithm to detect such cases. With the consideration of the detected pseudo multi-sense cases, we try to refine the existing word embeddings to eliminate the influence of pseudo multi-sense. Moreover, we apply our algorithm on previous released multi-sense word embeddings and tested it on artificial word similarity tasks and the analogy task. The result of the experiments shows that diminishing pseudo multi-sense can improve the quality of word representations. Thus, our method is actually an efficient way to reduce linguistic complexity.
In this paper, we present a comparative analysis of statistically predictive syntactic features of complexity and the treatment of these features by humans when simplifying texts. To that end, we have used a list of the most five statistically predictive features obtained automatically and the Corpus of Basque Simplified Texts (CBST) to analyse how the syntactic phenomena in these features have been manually simplified. Our aim is to go beyond the descriptions of operations found in the corpus and relate the multidisciplinary findings to understand text complexity from different points of view. We also present some issues that can be important when analysing linguistic complexity.
Pause analysis of key-stroke logged translations is a hallmark of process based translation studies. However, an exact definition of what a cognitively effortful pause during the translation process is has not been found yet (Saldanha and O’Brien, 2013). This paper investigates the design of a key-stroke and subject dependent identification system of cognitive effort to track complexity in translation with keystroke logging (cf. also (Dragsted, 2005) (Couto-Vale, in preparation)). It is an elastic measure that takes into account idiosyncratic pause duration of translators as well as further confounds such as bi-gram frequency, letter frequency and some motor tasks involved in writing. The method is compared to a common static threshold of 1000 ms in an analysis of cognitive effort during the translation of grammatical functions from English to German. Additionally, the results are triangulated with eye tracking data for further validation. The findings show that at least for smaller sets of data a dynamic pause assessment may lead to more accurate results than a generic static pause threshold of similar duration.
Data driven approaches to readability analysis for languages other than English has been plagued by a scarcity of suitable corpora. Often, relevant corpora consist only of easy-to-read texts with no rank information or empirical readability scores, making only binary approaches, such as classification, applicable. We propose a Bayesian, latent variable, approach to get the most out of these kinds of corpora. In this paper we present results on using such a model for readability ranking. The model is evaluated on a preliminary corpus of ranked student texts with encouraging results. We also assess the model by showing that it performs readability classification on par with a state of the art classifier while at the same being transparent enough to allow more sophisticated interpretations.
Informed by research on readability and language acquisition, computational linguists have developed sophisticated tools for the analysis of linguistic complexity. While some tools are starting to become accessible on the web, there still is a disconnect between the features that can in principle be identified based on state-of-the-art computational linguistic analysis, and the analyses a second language acquisition researcher, teacher, or textbook writer can readily obtain and visualize for their own collection of texts. This short paper presents a web-based tool development that aims to meet this challenge. The Common Text Analysis Platform (CTAP) is designed to support fully configurable linguistic feature extraction for a wide range of complexity analyses. It features a user-friendly interface, modularized and reusable analysis component integration, and flexible corpus and feature management. Building on the Unstructured Information Management framework (UIMA), CTAP readily supports integration of state-of-the-art NLP and complexity feature extraction maintaining modularization and reusability. CTAP thereby aims at providing a common platform for complexity analysis, encouraging research collaboration and sharing of feature extraction components—to jointly advance the state-of-the-art in complexity analysis in a form that readily supports real-life use by ordinary users.
We bring together knowledge from two different types of language learning data, texts learners read and texts they write, to improve linguistic complexity classification in the latter. Linguistic complexity in the foreign and second language learning context can be expressed in terms of proficiency levels. We show that incorporating features capturing lexical complexity information from reading passages can boost significantly the machine learning based classification of learner-written texts into proficiency levels. With an F1 score of .8 our system rivals state-of-the-art results reported for other languages for this task. Finally, we present a freely available web-based tool for proficiency level classification and lexical complexity visualization for both learner writings and reading texts.
Arabic writing is typically underspecified for short vowels and other markups, referred to as diacritics. In addition to the lexical ambiguity exhibited in most languages, the lack of diacritics in written Arabic adds another layer of ambiguity which is an artifact of the orthography. In this paper, we present the details of three annotation experimental conditions designed to study the impact of automatic ambiguity detection, on annotation speed and quality in a large scale annotation project.
Computational linguistic approaches to sign languages could benefit from investigating how complexity influences structure. We investigate whether morphological complexity has an effect on the order of Verb (V) and Object (O) in Swedish Sign Language (SSL), on the basis of elicited data from five Deaf signers. We find a significant difference in the distribution of the orderings OV vs. VO, based on an analysis of morphological weight. While morphologically heavy verbs exhibit a general preference for OV, humanness seems to affect the ordering in the opposite direction, with [+human] Objects pushing towards a preference for VO.
Language complexity is an intriguing phenomenon argued to play an important role in both language learning and processing. The need to compare languages with regard to their complexity resulted in a multitude of approaches and methods, ranging from accounts targeting specific structural features to global quantification of variation more generally. In this paper, we investigate the degree to which morphological complexity measures are mutually correlated in a sample of more than 500 languages of 101 language families. We use human expert judgements from the World Atlas of Language Structures (WALS), and compare them to four quantitative measures automatically calculated from language corpora. These consist of three previously defined corpus-derived measures, which are all monolingual, and one new measure based on automatic word-alignment across pairs of languages. We find strong correlations between all the measures, illustrating that both expert judgements and automated approaches converge to similar complexity ratings, and can be used interchangeably.
Comparable or parallel corpora are beneficial for many NLP tasks. The automatic collection of corpora enables large-scale resources, even for less-resourced languages, which in turn can be useful for deducing rules and patterns for text rewriting algorithms, a subtask of automatic text simplification. We present two methods for the alignment of Swedish easy-to-read text segments to text segments from a reference corpus. The first method (M1) was originally developed for the task of text reuse detection, measuring sentence similarity by a modified version of a TF-IDF vector space model. A second method (M2), also accounting for part-of-speech tags, was developed, and the methods were compared. For evaluation, a crowdsourcing platform was built for human judgement data collection, and preliminary results showed that cosine similarity relates better to human ranks than the Dice coefficient. We also saw a tendency that including syntactic context to the TF-IDF vector space model is beneficial for this kind of paraphrase alignment task.
This work presents a framework for the automatic construction of large Web corpora classified by readability level. We compare different Machine Learning classifiers for the task of readability assessment focusing on Portuguese and English texts, analysing the impact of variables like the feature inventory used in the resulting corpus. In a comparison between shallow and deeper features, the former already produce F-measures of over 0.75 for Portuguese texts, but the use of additional features results in even better results, in most cases. For English, shallow features also perform well as do classic readability formulas. Comparing different classifiers for the task, logistic regression obtained, in general, the best results, but with considerable differences between the results for two and those for three-classes, especially regarding the intermediary class. Given the large scale of the resulting corpus, for evaluation we adopt the agreement between different classifiers as an indication of readability assessment certainty. As a result of this work, a large corpus for Brazilian Portuguese was built, including 1.7 million documents and about 1.6 billion tokens, already parsed and annotated with 134 different textual attributes, along with the agreement among the various classifiers.
This work investigates the application of a measure of surprisal to modeling a grammatical variation phenomenon between near-synonymous constructions. We investigate a particular variation phenomenon, word order variation in Dutch two-verb clusters, where it has been established that word order choice is affected by processing cost. Several multifactorial corpus studies of Dutch verb clusters have used other measures of processing complexity to show that this factor affects word order choice. This previous work allows us to compare the surprisal measure, which is based on constraint satisfaction theories of language modeling, to those previously used measures, which are more directly linked to empirical observations of processing complexity. Our results show that surprisal does not predict the word order choice by itself, but is a significant predictor when used in a measure of uniform information density (UID). This lends support to the view that human language processing is facilitated not so much by predictable sequences of words but more by sequences of words in which information is spread evenly.
The relative contributions of meaning and form to sentence processing remains an outstanding issue across the language sciences. We examine this issue by formalizing four incremental complexity metrics and comparing them against freely-available ROI timecourses. Syntax-related metrics based on top-down parsing and structural dependency-distance turn out to significantly improve a regression model, compared to a simpler model that formalizes only conceptual combination using a distributional vector-space model. This confirms the view of the anterior temporal lobes as combinatory engines that deal in both form (see e.g. Brennan et al., 2012; Mazoyer, 1993) and meaning (see e.g., Patterson et al., 2007). This same characterization applies to a posterior temporal region in roughly “Wernicke’s Area.”
This paper investigates the use of automatic speech recognition (ASR) errors as indicators of the second language (L2) learners’ listening difficulties and in doing so strives to overcome the shortcomings of Partial and Synchronized Caption (PSC) system. PSC is a system that generates a partial caption including difficult words detected based on high speech rate, low frequency, and specificity. To improve the choice of words in this system, and explore a better method to detect speech challenges, ASR errors were investigated as a model of the L2 listener, hypothesizing that some of these errors are similar to those of language learners’ when transcribing the videos. To investigate this hypothesis, ASR errors in transcription of several TED talks were analyzed and compared with PSC’s selected words. Both the overlapping and mismatching cases were analyzed to investigate possible improvement for the PSC system. Those ASR errors that were not detected by PSC as cases of learners’ difficulties were further analyzed and classified into four categories: homophones, minimal pairs, breached boundaries and negatives. These errors were embedded into the baseline PSC to make the enhanced version and were evaluated in an experiment with L2 learners. The results indicated that the enhanced version, which encompasses the ASR errors addresses most of the L2 learners’ difficulties and better assists them in comprehending challenging video segments as compared with the baseline.
Eye-tracking reading times have been attested to reflect cognitive processes underlying sentence comprehension. However, the use of reading times in NLP applications is an underexplored area of research. In this initial work we build an automatic system to assess sentence complexity using automatically predicted eye-tracking reading time measures and demonstrate the efficacy of these reading times for a well known NLP task, namely, readability assessment. We use a machine learning model and a set of features known to be significant predictors of reading times in order to learn per-word reading times from a corpus of English text having reading times of human readers. Subsequently, we use the model to predict reading times for novel text in the context of the aforementioned task. A model based only on reading times gave competitive results compared to the systems that use extensive syntactic features to compute linguistic complexity. Our work, to the best of our knowledge, is the first study to show that automatically predicted reading times can successfully model the difficulty of a text and can be deployed in practical text processing applications.
The article presents results of entropy rate estimation for human languages across six languages by using large, state-of-the-art corpora of up to 7.8 gigabytes. To obtain the estimates for data length tending to infinity, we use an extrapolation function given by an ansatz. Whereas some ansatzes of this kind were proposed in previous research papers, here we introduce a stretched exponential extrapolation function that has a smaller error of fit. In this way, we uncover a possibility that the entropy rates of human languages are positive but 20% smaller than previously reported.
The morphological complexity of languages differs widely and changes over time. Pathways of change are often driven by the interplay of multiple competing factors, and are hard to disentangle. We here focus on a paradigmatic scenario of language change: the reduction of morphological complexity from Latin towards the Romance languages. To establish a causal explanation for this phenomenon, we employ three lines of evidence: 1) analyses of parallel corpora to measure the complexity of words in actual language production, 2) applications of NLP tools to further tease apart the contribution of inflectional morphology to word complexity, and 3) experimental data from artificial language learning, which illustrate the learning pressures at play when morphology simplifies. These three lines of evidence converge to show that pressures associated with imperfect language learning are good candidates to causally explain the reduction in morphological complexity in the Latin-to-Romance scenario. More generally, we argue that combining corpus, computational and experimental evidence is the way forward in historical linguistics and linguistic typology.
The goal of this paper is to examine the impact of simple feature engineering mechanisms before applying more sophisticated techniques to the task of medical NER. Sometimes papers using scientifically sound techniques present raw baselines that could be improved adding simple and cheap features. This work focuses on entity recognition for the clinical domain for three languages: English, Swedish and Spanish. The task is tackled using simple features, starting from the window size, capitalization, prefixes, and moving to POS and semantic tags. This work demonstrates that a simple initial step of feature engineering can improve the baseline results significantly. Hence, the contributions of this paper are: first, a short list of guidelines well supported with experimental results on three languages and, second, a detailed description of the relevance of these features for medical NER.
Automated extraction of concepts from patient clinical records is an essential facilitator of clinical research. For this reason, the 2010 i2b2/VA Natural Language Processing Challenges for Clinical Records introduced a concept extraction task aimed at identifying and classifying concepts into predefined categories (i.e., treatments, tests and problems). State-of-the-art concept extraction approaches heavily rely on handcrafted features and domain-specific resources which are hard to collect and define. For this reason, this paper proposes an alternative, streamlined approach: a recurrent neural network (the bidirectional LSTM with CRF decoding) initialized with general-purpose, off-the-shelf word embeddings. The experimental results achieved on the 2010 i2b2/VA reference corpora using the proposed framework outperform all recent methods and ranks closely to the best submission from the original 2010 i2b2/VA challenge.
Due to the recent replacements of physical documents with electronic medical records (EMR), the importance of information processing in medical fields has been increased. We have been organizing the MedNLP task series in NTCIR-10 and 11. These workshops were the first shared tasks which attempt to evaluate technologies that retrieve important information from medical reports written in Japanese. In this report, we describe the NTCIR-12 MedNLPDoc task which is designed for more advanced and practical use for the medical fields. This task is considered as a multi-labeling task to a patient record. This report presents results of the shared task, discusses and illustrates remained issues in the medical natural language processing field.
Patient notes contain a wealth of information of potentially great interest to medical investigators. However, to protect patients’ privacy, Protected Health Information (PHI) must be removed from the patient notes before they can be legally released, a process known as patient note de-identification. The main objective for a de-identification system is to have the highest possible recall. Recently, the first neural-network-based de-identification system has been proposed, yielding state-of-the-art results. Unlike other systems, it does not rely on human-engineered features, which allows it to be quickly deployed, but does not leverage knowledge from human experts or from electronic health records (EHRs). In this work, we explore a method to incorporate human-engineered features as well as features derived from EHRs to a neural-network-based de-identification system. Our results show that the addition of features, especially the EHR-derived features, further improves the state-of-the-art in patient note de-identification, including for some of the most sensitive PHI types such as patient names. Since in a real-life setting patient notes typically come with EHRs, we recommend developers of de-identification systems to leverage the information EHRs contain.
Semi-supervised clustering is an attractive alternative for traditional (unsupervised) clustering in targeted applications. By using the information of a small annotated dataset, semi-supervised clustering can produce clusters that are customized to the application domain. In this paper, we present a semi-supervised clustering technique based on a multi-objective evolutionary algorithm (NSGA-II-clus). We apply this technique to the task of clustering medical publications for Evidence Based Medicine (EBM) and observe an improvement of the results against unsupervised and other semi-supervised clustering techniques.
Rapid growth in Electronic Medical Records (EMR) has emerged to an expansion of data in the clinical domain. The majority of the available health care information is sealed in the form of narrative documents which form the rich source of clinical information. Text mining of such clinical records has gained huge attention in various medical applications like treatment and decision making. However, medical records enclose patient Private Health Information (PHI) which can reveal the identities of the patients. In order to retain the privacy of patients, it is mandatory to remove all the PHI information prior to making it publicly available. The aim is to de-identify or encrypt the PHI from the patient medical records. In this paper, we propose an algorithm based on deep learning architecture to solve this problem. We perform de-identification of seven PHI terms from the clinical records. Experiments on benchmark datasets show that our proposed approach achieves encouraging performance, which is better than the baseline model developed with Conditional Random Field.
Paraphrase generation is important in various applications such as search, summarization, and question answering due to its ability to generate textual alternatives while keeping the overall meaning intact. Clinical paraphrase generation is especially vital in building patient-centric clinical decision support (CDS) applications where users are able to understand complex clinical jargons via easily comprehensible alternative paraphrases. This paper presents Neural Clinical Paraphrase Generation (NCPG), a novel approach that casts the task as a monolingual neural machine translation (NMT) problem. We propose an end-to-end neural network built on an attention-based bidirectional Recurrent Neural Network (RNN) architecture with an encoder-decoder framework to perform the task. Conventional bilingual NMT models mostly rely on word-level modeling and are often limited by out-of-vocabulary (OOV) issues. In contrast, we represent the source and target paraphrase pairs as character sequences to address this limitation. To the best of our knowledge, this is the first work that uses attention-based RNNs for clinical paraphrase generation and also proposes an end-to-end character-level modeling for this task. Extensive experiments on a large curated clinical paraphrase corpus show that the attention-based NCPG models achieve improvements of up to 5.2 BLEU points and 0.5 METEOR points over a non-attention based strong baseline for word-level modeling, whereas further gains of up to 6.1 BLEU points and 1.3 METEOR points are obtained by the character-level NCPG models over their word-level counterparts. Overall, our models demonstrate comparable performance relative to the state-of-the-art phrase-based non-neural models.
The proliferation of deep learning methods in natural language processing (NLP) and the large amounts of data they often require stands in stark contrast to the relatively data-poor clinical NLP domain. In particular, large text corpora are necessary to build high-quality word embeddings, yet often large corpora that are suitably representative of the target clinical data are unavailable. This forces a choice between building embeddings from small clinical corpora and less representative, larger corpora. This paper explores this trade-off, as well as intermediate compromise solutions. Two standard clinical NLP tasks (the i2b2 2010 concept and assertion tasks) are evaluated with commonly used deep learning models (recurrent neural networks and convolutional neural networks) using a set of six corpora ranging from the target i2b2 data to large open-domain datasets. While combinations of corpora are generally found to work best, the single-best corpus is generally task-dependent.
Importance of utilizing medical information is getting increased as electronic health records (EHRs) are widely used nowadays. We aim to assign international standardized disease codes, ICD-10, to Japanese textual information in EHRs for users to reuse the information accurately. In this paper, we propose methods to automatically extract diagnosis and to assign ICD codes to Japanese medical records. Due to the lack of available training data, we dare employed rule-based methods rather than machine learning. We observed characteristics of medical records carefully, writing rules to make effective methods by hand. We applied our system to the NTCIR-12 MedNLPDoc shared task data where participants are required to assign ICD-10 codes of possible diagnosis in given EHRs. In this shared task, our system achieved the highest F-measure score among all participants in the most severe evaluation criteria. Through comparison with other approaches, we show that our approach could be a useful milestone for the future development of Japanese medical record processing.
In this work we present a fine-grained annotation schema to detect named entities in German clinical data of chronically ill patients with kidney diseases. The annotation schema is driven by the needs of our clinical partners and the linguistic aspects of German language. In order to generate annotations within a short period, the work also presents a semi-automatic annotation which uses additional sources of knowledge such as UMLS, to pre-annotate concepts in advance. The presented schema will be used to apply novel techniques from natural language processing and machine learning to support doctors treating their patients by improved information access from unstructured German texts.
In recent years, detecting Alzheimer disease (AD) in early stages based on natural language processing (NLP) has drawn much attention. To date, vocabulary size, grammatical complexity, and fluency have been studied using NLP metrics. However, the content analysis of AD narratives is still unreachable for NLP. This study investigates features of the words that AD patients use in their spoken language. After recruiting 18 examinees of 53–90 years old (mean: 76.89), they were divided into two groups based on MMSE scores. The AD group comprised 9 examinees with scores of 21 or lower. The healthy control group comprised 9 examinees with a score of 22 or higher. Linguistic Inquiry and Word Count (LIWC) classified words were used to categorize the words that the examinees used. The word frequency was found from observation. Significant differences were confirmed for the usage of impersonal pronouns in the AD group. This result demonstrated the basic feasibility of the proposed NLP-based detection approach.
The number of unstructured medical records kept in hospital information systems is increasing. The conditions of patients are formulated as outcomes in clinical pathway. A variance of an outcome describes deviations from standards of care like a patient’s bad condition. The present paper applied text mining to extract feature words and phrases of the variance from admission records. We report the cases the variances of “pain control” and “no neuropathy worsening” in cerebral infarction.
Clinical narratives in electronic health record systems are a rich resource of patient-based information. They constitute an ongoing challenge for natural language processing, due to their high compactness and abundance of short forms. German medical texts exhibit numerous ad-hoc abbreviations that terminate with a period character. The disambiguation of period characters is therefore an important task for sentence and abbreviation detection. This task is addressed by a combination of co-occurrence information of word types with trailing period characters, a large domain dictionary, and a simple rule engine, thus merging statistical and dictionary-based disambiguation strategies. An F-measure of 0.95 could be reached by using the unsupervised approach presented in this paper. The results are promising for a domain-independent abbreviation detection strategy, because our approach avoids retraining of models or use case specific feature engineering efforts required for supervised machine learning approaches.
The issue of privacy has always been a concern when clinical texts are used for research purposes. Personal health information (PHI) (such as name and identification number) needs to be removed so that patients cannot be identified. Manual anonymization is not feasible due to the large number of clinical texts to be anonymized. In this paper, we tackle the task of anonymizing clinical texts written in sentence fragments and which frequently contain symbols, abbreviations, and misspelled words. Our clinical texts therefore differ from those in the i2b2 shared tasks which are in prose form with complete sentences. Our clinical texts are also part of a structured database which contains patient name and identification number in structured fields. As such, we formulate our anonymization task as spelling variant detection, exploiting patients’ personal information in the structured fields to detect their spelling variants in clinical texts. We successfully anonymized clinical texts consisting of more than 200 million words, using minimum edit distance and regular expression patterns.
Men are from Mars and women are from Venus - or so the genre of relationship literature would have us believe. But there is some truth in this idea, and researchers in fields as diverse as psychology, sociology, and linguistics have explored ways to better understand the differences between genders. In this paper, we take another look at the problem of gender discrimination and attempt to move beyond the typical surface-level text classification approach, by (1) identifying semantic and psycholinguistic word classes that reflect systematic differences between men and women and (2) finding differences between genders in the ways they use the same words. We describe several experiments and report results on a large collection of blogs authored by men and women.
Recent studies have demonstrated gender and cultural differences in the recognition of emotions in facial expressions. However, most studies were conducted on American subjects. In this paper, we explore the generalizability of several findings to a non-American culture in the form of Danish subjects. We conduct an emotion recognition task followed by two stereotype questionnaires with different genders and age groups. While recent findings (Krems et al., 2015) suggest that women are biased to see anger in neutral facial expressions posed by females, in our sample both genders assign higher ratings of anger to all emotions expressed by females. Furthermore, we demonstrate an effect of gender on the fear-surprise-confusion observed by Tomkins and McCarter (1964); females overpredict fear, while males overpredict surprise.
Many methods have been used to recognise author personality traits from text, typically combining linguistic feature engineering with shallow learning models, e.g. linear regression or Support Vector Machines. This work uses deep-learning-based models and atomic features of text, the characters, to build hierarchical, vectorial word and sentence representations for trait inference. This method, applied to a corpus of tweets, shows state-of-the-art performance across five traits compared with prior work. The results, supported by preliminary visualisation work, are encouraging for the ability to detect complex human traits.
We exploit the Facebook reaction feature in a distant supervised fashion to train a support vector machine classifier for emotion detection, using several feature combinations and combining different Facebook pages. We test our models on existing benchmarks for emotion detection and show that employing only information that is derived completely automatically, thus without relying on any handcrafted lexicon as it’s usually done, we can achieve competitive results. The results also show that there is large room for improvement, especially by gearing the collection of Facebook pages, with a view to the target domain.
This paper proposes a graphical framework to extract opinionated sentences which highlight different contexts within a given news article by introducing the concept of diversity in a graphical model for opinion detection.We conduct extensive evaluations and find that the proposed modification leads to impressive improvement in performance and makes the final results of the model much more usable. The proposed method (OP-D) not only performs much better than the other techniques used for opinion detection as well as introducing diversity, but is also able to select opinions from different categories (Asher et al. 2009). By developing a classification model which categorizes the identified sentences into various opinion categories, we find that OP-D is able to push opinions from different categories uniformly among the top opinions.
Automatic detection of five language components, which are all relevant for expressing opinions and for stance taking, was studied: positive sentiment, negative sentiment, speculation, contrast and condition. A resource-aware approach was taken, which included manual annotation of 500 training samples and the use of limited lexical resources. Active learning was compared to random selection of training data, as well as to a lexicon-based method. Active learning was successful for the categories speculation, contrast and condition, but not for the two sentiment categories, for which results achieved when using active learning were similar to those achieved when applying a random selection of training data. This difference is likely due to a larger variation in how sentiment is expressed than in how speakers express the other three categories. This larger variation was also shown by the lower recall results achieved by the lexicon-based approach for sentiment than for the categories speculation, contrast and condition.
We investigate the application of kernel methods to representing both structural and lexical knowledge for predicting polarity of opinions in consumer product review. We introduce any-gram kernels which model lexical information in a significantly faster way than the traditional n-gram features, while capturing all possible orders of n-grams n in a sequence without the need to explicitly present a pre-specified set of such orders. We also present an effective format to represent constituency and dependency structure together with aspect terms and sentiment polarity scores. Furthermore, we modify the traditional tree kernel function to compute the similarity based on word embedding vectors instead of exact string match and present experiments using the new models.
This paper explores humour recognition for Twitter-based hashtag games. Given their popularity, frequency, and relatively formulaic nature, these games make a good target for computational humour research and can leverage Twitter likes and retweets as humour judgments. In this work, we use pair-wise relative humour judgments to examine several measures of semantic relatedness between setups and punchlines on a hashtag game corpus we collected and annotated. Results show that perplexity, Normalized Google Distance, and free-word association-based features are all useful in identifying “funnier” hashtag game responses. In fact, we provide empirical evidence that funnier punchlines tend to be more obscure, although more obscure punchlines are not necessarily rated funnier. Furthermore, the asymmetric nature of free-word association features allows us to see that while punchlines should be harder to predict given a setup, they should also be relatively easy to understand in context.
Despite a substantial progress made in developing new sentiment lexicon generation (SLG) methods for English, the task of transferring these approaches to other languages and domains in a sound way still remains open. In this paper, we contribute to the solution of this problem by systematically comparing semi-automatic translations of common English polarity lists with the results of the original automatic SLG algorithms, which were applied directly to German data. We evaluate these lexicons on a corpus of 7,992 manually annotated tweets. In addition to that, we also collate the results of dictionary- and corpus-based SLG methods in order to find out which of these paradigms is better suited for the inherently noisy domain of social media. Our experiments show that semi-automatic translations notably outperform automatic systems (reaching a macro-averaged F1-score of 0.589), and that dictionary-based techniques produce much better polarity lists as compared to corpus-based approaches (whose best F1-scores run up to 0.479 and 0.419 respectively) even for the non-standard Twitter genre.
Detecting depression or personality traits, tutoring and student behaviour systems, or identifying cases of cyber-bulling are a few of the wide range of the applications, in which the automatic detection of emotion is a crucial element. Emotion detection has the potential of high impact by contributing the benefit of business, society, politics or education. Given this context, the main objective of our research is to contribute to the resolution of one of the most important challenges in textual emotion detection task: the problems of emotional corpora annotation. This will be tackled by proposing of a new semi-automatic methodology. Our innovative methodology consists in two main phases: (1) an automatic process to pre-annotate the unlabelled sentences with a reduced number of emotional categories; and (2) a refinement manual process where human annotators will determine which is the predominant emotion between the emotional categories selected in the phase 1. Our proposal in this paper is to show and evaluate the pre-annotation process to analyse the feasibility and the benefits by the methodology proposed. The results obtained are promising and allow obtaining a substantial improvement of annotation time and cost and confirm the usefulness of our pre-annotation process to improve the annotation task.
We created a model to estimate personality trait from authors’ text written in Japanese and measured its performance by conducting surveys and analyzing the Twitter data of 1,630 users. We used the Big Five personality traits for personality trait estimation. Our approach is a combination of category- and Word2Vec-based approaches. For the category-based element, we added several unique Japanese categories along with the ones regularly used in the English model, and for the Word2Vec-based element, we used a model called GloVe. We found that some of the newly added categories have a stronger correlation with personality traits than other categories do and that the combination of the category- and Word2Vec-based approaches improves the accuracy of the personality trait estimation compared with the case of using just one of them.
On June 23rd 2016, UK held the referendum which ratified the exit from the EU. While most of the traditional pollsters failed to forecast the final vote, there were online systems that hit the result with high accuracy using opinion mining techniques and big data. Starting one month before, we collected and monitored millions of posts about the referendum from social media conversations, and exploited Natural Language Processing techniques to predict the referendum outcome. In this paper we discuss the methods used by traditional pollsters and compare it to the predictions based on different opinion mining techniques. We find that opinion mining based on agreement/disagreement classification works better than opinion mining based on polarity classification in the forecast of the referendum outcome.
The conundrum of understanding and classifying sarcasm has been dealt with by the traditional theorists as an analysis of a sarcastic utterance and the ironic situation that surrounds it. The problem with such an approach is that it is too narrow, as it is unable to sufficiently utilize the two indispensable agents in making such an utterance, viz. the speaker and the listener. It undermines the necessary context required to comprehend a sarcastic utterance. In this paper, we propose a novel approach towards understanding sarcasm in terms of the existing knowledge hierarchy between the two participants, which forms the basis of the context that both agents share. The difference in relationship of the speaker of the sarcastic utterance and the disparate audience found on social media, such as Twitter, is also captured. We then apply our model on a corpus of tweets to achieve significant results and consequently, shed light on subjective nature of context, which is contingent on the relation between the speaker and the listener.
A growing body of research exploits social media behaviors to gauge psychological character-istics, though trait empathy has received little attention. Because of its intimate link to the abil-ity to relate to others, our research aims to predict participants’ levels of empathy, given their textual and friending behaviors on Facebook. Using Poisson regression, we compared the vari-ance explained in Davis’ Interpersonal Reactivity Index (IRI) scores on four constructs (em-pathic concern, personal distress, fantasy, perspective taking), by two classes of variables: 1) post content and 2) linguistic style. Our study lays the groundwork for a greater understanding of empathy’s role in facilitating interactions on social media.
This paper outlines a pilot study on multi-dimensional and multilingual sentiment analysis of social media content. We use parallel corpora of movie subtitles as a proxy for colloquial language in social media channels and a multilingual emotion lexicon for fine-grained sentiment analyses. Parallel data sets make it possible to study the preservation of sentiments and emotions in translation and our assessment reveals that the lexical approach shows great inter-language agreement. However, our manual evaluation also suggests that the use of purely lexical methods is limited and further studies are necessary to pinpoint the cross-lingual differences and to develop better sentiment classifiers.
In this paper, we address the issue of automatic prediction of readers’ mood from newspaper articles and comments. As online newspapers are becoming more and more similar to social media platforms, users can provide affective feedback, such as mood and emotion. We have exploited the self-reported annotation of mood categories obtained from the metadata of the Italian online newspaper corriere.it to design and evaluate a system for predicting five different mood categories from news articles and comments: indignation, disappointment, worry, satisfaction, and amusement. The outcome of our experiments shows that overall, bag-of-word-ngrams perform better compared to all other feature sets; however, stylometric features perform better for the mood score prediction of articles. Our study shows that self-reported annotations can be used to design automatic mood prediction systems.
Most work in NLP analysing microblogs focuses on textual content thus neglecting temporal and spatial information. We present a new interdisciplinary method for emotion classification that combines linguistic, temporal, and spatial information into a single metric. We create a graph of labeled and unlabeled tweets that encodes the relations between neighboring tweets with respect to their emotion labels. Graph-based semi-supervised learning labels all tweets with an emotion.
We explore a domain-agnostic approach for analyzing speech with the goal of opinion prediction. We represent the speech signal by mel-frequency cepstral coefficients and apply long short-term memory neural networks to automatically learn temporal regularities in speech. In contrast to previous work, our approach does not require complex feature engineering and works without textual transcripts. As a consequence, it can easily be applied on various speech analysis tasks for different languages and the results show that it can nevertheless be competitive to the state-of-the-art in opinion prediction. In a detailed error analysis for opinion mining we find that our approach performs well in identifying speaker-specific characteristics, but should be combined with additional information if subtle differences in the linguistic content need to be identified.
Considering the importance of public speech skills, a system which makes a prediction on where audiences laugh in a talk can be helpful to a person who prepares for a talk. We investigated a possibility that a state-of-the-art humor recognition system can be used in detecting sentences inducing laughters in talks. In this study, we used TED talks and laughters in the talks as data. Our results showed that the state-of-the-art system needs to be improved in order to be used in a practical application. In addition, our analysis showed that classifying humorous sentences in talks is very challenging due to close distance between humorous and non-humorous sentences.
Major depressive disorder, a debilitating and burdensome disease experienced by individuals worldwide, can be defined by several depressive symptoms (e.g., anhedonia (inability to feel pleasure), depressed mood, difficulty concentrating, etc.). Individuals often discuss their experiences with depression symptoms on public social media platforms like Twitter, providing a potentially useful data source for monitoring population-level mental health risk factors. In a step towards developing an automated method to estimate the prevalence of symptoms associated with major depressive disorder over time in the United States using Twitter, we developed classifiers for discerning whether a Twitter tweet represents no evidence of depression or evidence of depression. If there was evidence of depression, we then classified whether the tweet contained a depressive symptom and if so, which of three subtypes: depressed mood, disturbed sleep, or fatigue or loss of energy. We observed that the most accurate classifiers could predict classes with high-to-moderate F1-score performances for no evidence of depression (85), evidence of depression (52), and depressive symptoms (49). We report moderate F1-scores for depressive symptoms ranging from 75 (fatigue or loss of energy) to 43 (disturbed sleep) to 35 (depressed mood). Our work demonstrates baseline approaches for automatically encoding Twitter data with granular depressive symptoms associated with major depressive disorder.
This paper proposes a new idea that uses Wikipedia categories as answer types and defines candidate sets inside Wikipedia. The focus of a given question is searched in the hierarchy of Wikipedia main pages. Our searching strategy combines head-noun matching and synonym matching provided in semantic resources. The set of answer candidates is determined by the entry hierarchy in Wikipedia and the hyponymy hierarchy in WordNet. The experimental results show that the approach can find candidate sets in a smaller size but achieve better performance especially for ARTIFACT and ORGANIZATION types, where the performance is better than state-of-the-art Chinese factoid QA systems.
Commonsense knowledge is essential for fully understanding language in many situations. We acquire large-scale commonsense knowledge from humans using a game with a purpose (GWAP) developed on a smartphone spoken dialogue system. We transform the manual knowledge acquisition process into an enjoyable quiz game and have collected over 150,000 unique commonsense facts by gathering the data of more than 70,000 players over eight months. In this paper, we present a simple method for maintaining the quality of acquired knowledge and an empirical analysis of the knowledge acquisition process. To the best of our knowledge, this is the first work to collect large-scale knowledge via a GWAP on a widely-used spoken dialogue system.
This paper describes a hierarchical neural network we propose for sentence classification to extract product information from product documents. The network classifies each sentence in a document into attribute and condition classes on the basis of word sequences and sentence sequences in the document. Experimental results showed the method using the proposed network significantly outperformed baseline methods by taking semantic representation of word and sentence sequential data into account. We also evaluated the network with two different product domains (insurance and tourism domains) and found that it was effective for both the domains.
Question answering is always an attractive and challenging task in natural language processing area. There are some open domain question answering systems, such as IBM Waston, which take the unstructured text data as input, in some ways of humanlike thinking process and a mode of artificial intelligence. At the conference on Natural Language Processing and Chinese Computing (NLPCC) 2016, China Computer Federation hosted a shared task evaluation about Open Domain Question Answering. We achieve the 2nd place at the document-based subtask. In this paper, we present our solution, which consists of feature engineering in lexical and semantic aspects and model training methods. As the result of the evaluation shows, our solution provides a valuable and brief model which could be used in modelling question answering or sentence semantic relevance. We hope our solution would contribute to this vast and significant task with some heuristic thinking.
An Entity-based approach to Answering recurrent and non-recurrent questions with Past Answers Abstract Community question answering (CQA) systems such as Yahoo! Answers allow registered-users to ask and answer questions in various question categories. However, a significant percentage of asked questions in Yahoo! Answers are unanswered. In this paper, we propose to reduce this percentage by reusing answers to past resolved questions from the site. Specifically, we propose to satisfy unanswered questions in entity rich categories by searching for and reusing the best answers to past resolved questions with shared needs. For unanswered questions that do not have a past resolved question with a shared need, we propose to use the best answer to a past resolved question with similar needs. Our experiments on a Yahoo! Answers dataset shows that our approach retrieves most of the past resolved questions that have shared and similar needs to unanswered questions.
In an era where highly accurate Question Answering (QA) systems are being built using complex Natural Language Processing (NLP) and Information Retrieval (IR) algorithms, presenting the acquired answer to the user akin to a human answer is also crucial. In this paper we present an answer presentation strategy by embedding the answer in a sentence which is developed by incorporating the linguistic structure of the source question extracted through typed dependency parsing. The evaluation using human participants proved that the methodology is human-competitive and can result in linguistically correct sentences for more that 70% of the test dataset acquired from QALD question dataset.
Question answering (QA) systems need to provide exact answers for the questions that are posed to the system. However, this can only be achieved through a precise processing of the question. During this procedure, one important step is the detection of the expected type of answer that the system should provide by extracting the headword of the questions and identifying its semantic type. We have annotated the headword and assigned UMLS semantic types to 643 factoid/list questions from the BioASQ training data. We present statistics on the corpus and a preliminary evaluation in baseline experiments. We also discuss the challenges on both the manual annotation and the automatic detection of the headwords and the semantic types. We believe that this is a valuable resource for both training and evaluation of biomedical QA systems. The corpus is available at: https://github.com/mariananeves/BioMedLAT.
The paper describes topic shifting in dialogues with a robot that provides information from Wiki-pedia. The work focuses on a double topical construction of dialogue coherence which refers to discourse coherence on two levels: the evolution of dialogue topics via the interaction between the user and the robot system, and the creation of discourse topics via the content of the Wiki-pedia article itself. The user selects topics that are of interest to her, and the system builds a list of potential topics, anticipated to be the next topic, by the links in the article and by the keywords extracted from the article. The described system deals with Wikipedia articles, but could easily be adapted to other digital information providing systems.
Wikipedia has become a reference knowledge source for scores of NLP applications. One of its invaluable features lies in its multilingual nature, where articles on a same entity or concept can have from one to more than 200 different versions. The interlinking of language versions in Wikipedia has undergone a major renewal with the advent of Wikidata, a unified scheme to identify entities and their properties using unique numbers. However, as the interlinking is still manually carried out by thousands of editors across the globe, errors may creep in the assignment of entities. In this paper, we describe an optimization technique to match automatically language versions of articles, and hence entities, that is only based on bags of words and anchors. We created a dataset of all the articles on persons we extracted from Wikipedia in six languages: English, French, German, Russian, Spanish, and Swedish. We report a correct match of at least 94.3% on each pair.
In this paper, we present an open information extraction system so-called SRDF that generates lexical knowledge graphs from unstructured texts. In semantic web, knowledge is expressed in the RDF triple form but the natural language text consist of multiple relations between arguments. For this reason, we combine open information extraction with the reification for the full text extraction to preserve meaning of sentence in our knowledge graph. And also our knowledge graph is designed to adapt for many existing semantic web applications. At the end of this paper, we introduce the result of the experiment and a Korean template generation module developed using SRDF.
Natural language questions are interpreted to a sequence of patterns to be matched with instances of patterns in a knowledge base (KB) for answering. A natural language (NL) question answering (QA) system utilizes meaningful patterns matching the syntac-tic/lexical features between the NL questions and KB. In the most of KBs, there are only binary relations in triple form to represent relation between two entities or entity and a value using the domain specific ontology. However, the binary relation representation is not enough to cover complex information in questions, and the ontology vocabulary sometimes does not cover the lexical meaning in questions. Complex meaning needs a knowledge representation to link the binary relation-type triples in KB. In this paper, we propose a frame semantics-based semantic parsing approach as KB-independent question pre-processing. We will propose requirements of question interpretation in the KBQA perspective, and a query form representation based on our proposed format QAF (Ques-tion Answering with the Frame Semantics), which is supposed to cover the requirements. In QAF, frame semantics roles as a model to represent complex information in questions and to disambiguate the lexical meaning in questions to match with the ontology vocabu-lary. Our system takes a question as an input and outputs QAF-query by the process which assigns semantic information in the question to its corresponding frame semantic structure using the semantic parsing rules.
Answering yes–no questions is more difficult than simply retrieving ranked search results. To answer yes–no questions, especially when the correct answer is no, one must find an objectionable keyword that makes the question’s answer no. Existing systems, such as factoid-based ones, cannot answer yes–no questions very well because of insufficient handling of such objectionable keywords. We suggest an algorithm that answers yes–no questions by assigning an importance to objectionable keywords. Concretely speaking, we suggest a penalized scoring method that finds and makes lower score for parts of documents that include such objectionable keywords. We check a keyword distribution for each part of a document such as a paragraph, calculating the keyword density as a basic score. Then we use an objectionable keyword penalty when a keyword does not appear in a target part but appears in other parts of the document. Our algorithm is robust for open domain problems because it requires no training. We achieved 4.45 point better results in F1 scores than the best score of the NTCIR-10 RITE2 shared task, also obtained the best score in 2014 mock university examination challenge of the Todai Robot project.
Nowadays, a question answering (QA) system is used in various areas such a quiz show, personal assistant, home device, and so on. The OKBQA framework supports developing a QA system in an intuitive and collaborative ways. To support collaborative development, the framework should be equipped with some functions, e.g., flexible system configuration, debugging supports, intuitive user interface, and so on while considering different developing groups of different domains. This paper presents OKBQA controller, a dedicated workflow manager for OKBQA framework, to boost collaborative development of a QA system.
fast align is a simple and fast word alignment tool which is widely used in state-of-the-art machine translation systems. It yields comparable results in the end-to-end translation experiments of various language pairs. However, fast align does not perform as well as GIZA++ when applied to language pairs with distinct word orders, like English and Japanese. In this paper, given the lexical translation table output by fast align, we propose to realign words using the hierarchical sub-sentential alignment approach. Experimental results show that simple additional processing improves the performance of word alignment, which is measured by counting alignment matches in comparison with fast align. We also report the result of final machine translation in both English-Japanese and Japanese-English. We show our best system provided significant improvements over the baseline as measured by BLEU and RIBES.
This paper presents the comparison of how using different neural network based language modeling tools for selecting the best candidate fragments affects the final output translation quality in a hybrid multi-system machine translation setup. Experiments were conducted by comparing perplexity and BLEU scores on common test cases using the same training data set. A 12-gram statistical language model was selected as a baseline to oppose three neural network based models of different characteristics. The models were integrated in a hybrid system that depends on the perplexity score of a sentence fragment to produce the best fitting translations. The results show a correlation between language model perplexity and BLEU scores as well as overall improvements in BLEU.
We present a novel method of comparable corpora construction. Unlike the traditional methods which heavily rely on linguistic features, our method only takes image similarity into consid-eration. We use an image-image search engine to obtain similar images, together with the cap-tions in source language and target language. On the basis, we utilize captions of similar imag-es to construct sentence-level bilingual corpora. Experiments on 10,371 target captions show that our method achieves a precision of 0.85 in the top search results.
We present an algorithm for predicting translation equivalents between two languages, based on the corresponding WordNets. The assumption is that all synsets of one of the languages are linked to the corresponding synsets in the other language. In theory, given the exact sense of a word in a context it must be possible to translate it as any of the words in the linked synset. In practice, however, this does not work well since automatic and accurate sense disambiguation is difficult. Instead it is possible to define a more robust translation relation between the lexemes of the two languages. As far as we know the Finnish WordNet is the only one that includes that relation. Our algorithm can be used to predict the relation for other languages as well. This is useful for instance in hybrid machine translation systems which are usually more dependent on high-quality translation dictionaries.
Traditional machine translation evaluation metrics such as BLEU and WER have been widely used, but these metrics have poor correlations with human judgements because they badly represent word similarity and impose strict identity matching. In this paper, we propose some modifications to the traditional measures based on word embeddings for these two metrics. The evaluation results show that our modifications significantly improve their correlation with human judgements.
We describe experiments in Machine Translation using word sense disambiguation (WSD) information. This work focuses on WSD in verbs, based on two different approaches – verbal patterns based on corpus pattern analysis and verbal word senses from valency frames. We evaluate several options of using verb senses in the source-language sentences as an additional factor for the Moses statistical machine translation system. Our results show a statistically significant translation quality improvement in terms of the BLEU metric for the valency frames approach, but in manual evaluation, both WSD methods bring improvements.
We introduce a new statistical machine translation approach specifically geared to learning translation from low resource languages, that exploits monolingual English semantic parsing to bias inversion transduction grammar (ITG) induction. We show that in contrast to conventional statistical machine translation (SMT) training methods, which rely heavily on phrase memorization, our approach focuses on learning bilingual correlations that help translating low resource languages, by using the output language semantic structure to further narrow down ITG constraints. This approach is motivated by previous research which has shown that injecting a semantic frame based objective function while training SMT models improves the translation quality. We show that including a monolingual semantic objective function during the learning of the translation model leads towards a semantically driven alignment which is more efficient than simply tuning loglinear mixture weights against a semantic frame based evaluation metric in the final stage of statistical machine translation training. We test our approach with three different language pairs and demonstrate that our model biases the learning towards more semantically correct alignments. Both GIZA++ and ITG based techniques fail to capture meaningful bilingual constituents, which is required when trying to learn translation models for low resource languages. In contrast, our proposed model not only improve translation by injecting a monolingual objective function to learn bilingual correlations during early training of the translation model, but also helps to learn more meaningful correlations with a relatively small data set, leading to a better alignment compared to either conventional ITG or traditional GIZA++ based approaches.
We defend that bilingual lexicons automatically extracted from parallel corpora, whose entries have been meanwhile validated by linguists and classified as correct or incorrect, should constitute a specific parallel corpora. And, in this paper, we propose to use word-to-word translations to learn morph-units (comprising of bilingual stems and suffixes) from those bilingual lexicons for two language pairs L1-L2 and L1-L3 to induce a bilingual lexicon for the language pair L2-L3, apart from also learning morph-units for this other language pair. The applicability of bilingual morph-units in L1-L2 and L1-L3 is examined from the perspective of pivot-based lexicon induction for language pair L2-L3 with L1 as bridge. While the lexicon is derived by transitivity, the correspondences are identified based on previously learnt bilingual stems and suffixes rather than surface translation forms. The induced pairs are validated using a binary classifier trained on morphological and similarity-based features using an existing, automatically acquired, manually validated bilingual translation lexicon for language pair L2-L3. In this paper, we discuss the use of English (EN)-French (FR) and English (EN)-Portuguese (PT) lexicon of word-to-word translations in generating word-to-word translations for the language pair FR-PT with EN as pivot language. Generated translations are filtered out first using an SVM-based FR-PT classifier and then are manually validated.
This paper presents the results of the shared tasks from the 3rd workshop on Asian translation (WAT2016) including J ↔ E, J ↔ C scientific paper translation subtasks, C ↔ J, K ↔ J, E ↔ J patent translation subtasks, I ↔ E newswire subtasks and H ↔ E, H ↔ J mixed domain subtasks. For the WAT2016, 15 institutions participated in the shared tasks. About 500 translation results have been submitted to the automatic evaluation server, and selected submissions were manually evaluated.
Neural machine translation (NMT), a new approach to machine translation, has achieved promising results comparable to those of traditional approaches such as statistical machine translation (SMT). Despite its recent success, NMT cannot handle a larger vocabulary because training complexity and decoding complexity proportionally increase with the number of target words. This problem becomes even more serious when translating patent documents, which contain many technical terms that are observed infrequently. In NMTs, words that are out of vocabulary are represented by a single unknown token. In this paper, we propose a method that enables NMT to translate patent sentences comprising a large vocabulary of technical terms. We train an NMT system on bilingual data wherein technical terms are replaced with technical term tokens; this allows it to translate most of the source sentences except technical terms. Further, we use it as a decoder to translate source sentences with technical term tokens and replace the tokens with technical term translations using SMT. We also use it to rerank the 1,000-best SMT translations on the basis of the average of the SMT score and that of the NMT rescoring of the translated sentences with technical term tokens. Our experiments on Japanese-Chinese patent sentences show that the proposed NMT system achieves a substantial improvement of up to 3.1 BLEU points and 2.3 RIBES points over traditional SMT systems and an improvement of approximately 0.6 BLEU points and 0.8 RIBES points over an equivalent NMT system without our proposed technique.
Concomitant with the globalization of food culture, demand for the recipes of specialty dishes has been increasing. The recent growth in recipe sharing websites and food blogs has resulted in numerous recipe texts being available for diverse foods in various languages. However, little work has been done on machine translation of recipe texts. In this paper, we address the task of translating recipes and investigate the advantages and disadvantages of traditional phrase-based statistical machine translation and more recent neural machine translation. Specifically, we translate Japanese recipes into English, analyze errors in the translated recipes, and discuss available room for improvements.
This paper describes the IIT Bombay’s submission as a part of the shared task in WAT 2016 for English–Indonesian language pair. The results reported here are for both the direction of the language pair. Among the various approaches experimented, Operation Sequence Model (OSM) and Neural Language Model have been submitted for WAT. The OSM approach integrates translation and reordering process resulting in relatively improved translation. Similarly the neural experiment integrates Neural Language Model with Statistical Machine Translation (SMT) as a feature for translation. The Neural Probabilistic Language Model (NPLM) gave relatively high BLEU points for Indonesian to English translation system while the Neural Network Joint Model (NNJM) performed better for English to Indonesian direction of translation system. The results indicate improvement over the baseline Phrase-based SMT by 0.61 BLEU points for English-Indonesian system and 0.55 BLEU points for Indonesian-English translation system.
This paper describes our UT-KAY system that participated in the Workshop on Asian Translation 2016. Based on an Attention-based Neural Machine Translation (ANMT) model, we build our system by incorporating a domain adaptation method for multiple domains and an attention-based unknown word replacement method. In experiments, we verify that the attention-based unknown word replacement method is effective in improving translation scores in Chinese-to-Japanese machine translation. We further show results of manual analysis on the replaced unknown words.
When translating formal documents, capturing the sentence structure specific to the sublanguage is extremely necessary to obtain high-quality translations. This paper proposes a novel global reordering method with particular focus on long-distance reordering for capturing the global sentence structure of a sublanguage. The proposed method learns global reordering models from a non-annotated parallel corpus and works in conjunction with conventional syntactic reordering. Experimental results on the patent abstract sublanguage show substantial gains of more than 25 points in the RIBES metric and comparable BLEU scores both for Japanese-to-English and English-to-Japanese translations.
This paper presents an improved lexicalized reordering model for phrase-based statistical machine translation using a deep neural network. Lexicalized reordering suffers from reordering ambiguity, data sparseness and noises in a phrase table. Previous neural reordering model is successful to solve the first and second problems but fails to address the third one. Therefore, we propose new features using phrase translation and word alignment to construct phrase vectors to handle inherently noisy phrase translation pairs. The experimental results show that our proposed method improves the accuracy of phrase reordering. We confirm that the proposed method works well with phrase pairs including NULL alignments.
This paper presents our machine translation system that developed for the WAT2016 evalua-tion tasks of ja-en, ja-zh, en-ja, zh-ja, JPCja-en, JPCja-zh, JPCen-ja, JPCzh-ja. We build our system based on encoder–decoder framework by integrating recurrent neural network (RNN) and gate recurrent unit (GRU), and we also adopt an attention mechanism for solving the problem of information loss. Additionally, we propose a simple translation-specific approach to resolve the unknown word translation problem. Experimental results show that our system performs better than the baseline statistical machine translation (SMT) systems in each task. Moreover, it shows that our proposed approach of unknown word translation performs effec-tively improvement of translation results.
System architecture, experimental settings and experimental results of the group for the WAT2016 tasks are described. We participate in six tasks: en-ja, zh-ja, JPCzh-ja, JPCko-ja, HINDENen-hi and HINDENhi-ja. Although the basic architecture of our sys-tems is PBSMT with reordering, several techniques are conducted. Especially, the system for the HINDENhi-ja task with pivoting by English uses the reordering technique. Be-cause Hindi and Japanese are both OV type languages and English is a VO type language, we can use reordering technique to the pivot language. We can improve BLEU score from 7.47 to 7.66 by the reordering technique for the sentence level pivoting of this task.
This year, the Nara Institute of Science and Technology (NAIST)/Carnegie Mellon University (CMU) submission to the Japanese-English translation track of the 2016 Workshop on Asian Translation was based on attentional neural machine translation (NMT) models. In addition to the standard NMT model, we make a number of improvements, most notably the use of discrete translation lexicons to improve probability estimates, and the use of minimum risk training to optimize the MT system for BLEU score. As a result, our system achieved the highest translation evaluation scores for the task.
This paper describes the NICT-2 translation system for the 3rd Workshop on Asian Translation. The proposed system employs a domain adaptation method based on feature augmentation. We regarded the Japan Patent Office Corpus as a mixture of four domain corpora and improved the translation quality of each domain. In addition, we incorporated language models constructed from Google n-grams as external knowledge. Our domain adaptation method can naturally incorporate such external knowledge that contributes to translation quality.
We participate in scientific paper subtask (ASPEC-EJ/CJ) and patent subtask (JPC-EJ/CJ/KJ) with phrase-based SMT systems which are trained with its own patent corpora. Using larger corpora than those prepared by the workshop organizer, we achieved higher BLEU scores than most participants in EJ and CJ translations of patent subtask, but in crowdsourcing evaluation, our EJ translation, which is best in all automatic evaluations, received a very poor score. In scientific paper subtask, our translations are given lower scores than most translations that are produced by translation engines trained with the in-domain corpora. But our scores are higher than those of general-purpose RBMTs and online services. Considering the result of crowdsourcing evaluation, it shows a possibility that CJ SMT system trained with a large patent corpus translates non-patent technical documents at a practical level.
Simultaneous interpretation is a very challenging application of machine translation in which the input is a stream of words from a speech recognition engine. The key problem is how to segment the stream in an online manner into units suitable for translation. The segmentation process proceeds by calculating a confidence score for each word that indicates the soundness of placing a sentence boundary after it, and then heuristics are employed to determine the position of the boundaries. Multiple variants of the confidence scoring method and segmentation heuristics were studied. Experimental results show that the best performing strategy is not only efficient in terms of average latency per word, but also achieved end-to-end translation quality close to an offline baseline, and close to oracle segmentation.
This paper illustrates the similarity between Thai and Laotian, and between Malay and Indonesian, based on an investigation on raw parallel data from Asian Language Treebank. The cross-lingual similarity is investigated and demonstrated on metrics of correspondence and order of tokens, based on several standard statistical machine translation techniques. The similarity shown in this study suggests a possibility on harmonious annotation and processing of the language pairs in future development.
We propose a method for integrating Japanese empty category detection into the preordering process of Japanese-to-English statistical machine translation. First, we apply machine-learning-based empty category detection to estimate the position and the type of empty categories in the constituent tree of the source sentence. Then, we apply discriminative preordering to the augmented constituent tree in which empty categories are treated as if they are normal lexical symbols. We find that it is effective to filter empty categories based on the confidence of estimation. Our experiments show that, for the IWSLT dataset consisting of short travel conversations, the insertion of empty categories alone improves the BLEU score from 33.2 to 34.3 and the RIBES score from 76.3 to 78.7, which imply that reordering has improved For the KFTT dataset consisting of Wikipedia sentences, the proposed preordering method considering empty categories improves the BLEU score from 19.9 to 20.2 and the RIBES score from 66.2 to 66.3, which shows both translation and reordering have improved slightly.
We describe here our approaches and results on the WAT 2016 shared translation tasks. We tried to use both an example-based machine translation (MT) system and a neural MT system. We report very good translation results, especially when using neural MT for Chinese-to-Japanese translation.
This paper reports our systems (UT-AKY) submitted in the 3rd Workshop of Asian Translation 2016 (WAT’16) and their results in the English-to-Japanese translation task. Our model is based on the tree-to-sequence Attention-based NMT (ANMT) model proposed by Eriguchi et al. (2016). We submitted two ANMT systems: one with a word-based decoder and the other with a character-based decoder. Experimenting on the English-to-Japanese translation task, we have confirmed that the character-based decoder can cover almost the full vocabulary in the target language and generate translations much faster than the word-based model.
This paper describes the SENSE machine translation system participation in the Third Workshop for Asian Translation (WAT2016). We share our best practices to build a fast and light phrase-based machine translation (PBMT) models that have comparable results to the baseline systems provided by the organizers. As Neural Machine Translation (NMT) overtakes PBMT as the state-of-the-art, deep learning and new MT practitioners might not be familiar with the PBMT paradigm and we hope that this paper will help them build a PBMT baseline system quickly and easily.
Unlike European languages, many Asian languages like Chinese and Japanese do not have typographic boundaries in written system. Word segmentation (tokenization) that break sentences down into individual words (tokens) is normally treated as the first step for machine translation (MT). For Chinese and Japanese, different rules and segmentation tools lead different segmentation results in different level of granularity between Chinese and Japanese. To improve the translation accuracy, we adjust and balance the granularity of segmentation results around terms for Chinese–Japanese patent corpus for training translation model. In this paper, we describe a statistical machine translation (SMT) system which is built on re-tokenized Chinese-Japanese patent training corpus using extracted bilingual multi-word terms.
In machine translation, we must consider the difference in expression between languages. For example, the active/passive voice may change in Japanese-English translation. The same verb in Japanese may be translated into different voices at each translation because the voice of a generated sentence cannot be determined using only the information of the Japanese sentence. Machine translation systems should consider the information structure to improve the coherence of the output by using several topicalization techniques such as passivization. Therefore, this paper reports on our attempt to control the voice of the sentence generated by an encoder-decoder model. To control the voice of the generated sentence, we added the voice information of the target sentence to the source sentence during the training. We then generated sentences with a specified voice by appending the voice information to the source sentence. We observed experimentally whether the voice could be controlled. The results showed that, we could control the voice of the generated sentence with 85.0% accuracy on average. In the evaluation of Japanese-English translation, we obtained a 0.73-point improvement in BLEU score by using gold voice labels.
This paper presents our Chinese-to-Japanese patent machine translation system for WAT 2016 (Group ID: ntt) that uses syntactic pre-ordering over Chinese dependency structures. Chinese words are reordered by a learning-to-rank model based on pairwise classification to obtain word order close to Japanese. In this year’s system, two different machine translation methods are compared: traditional phrase-based statistical machine translation and recent sequence-to-sequence neural machine translation with an attention mechanism. Our pre-ordering showed a significant improvement over the phrase-based baseline, but, in contrast, it degraded the neural machine translation baseline.
In this paper we describe the system that we develop as part of our participation in WAT 2016. We develop a system based on hierarchical phrase-based SMT for English to Hindi language pair. We perform re-ordering and augment bilingual dictionary to improve the performance. As a baseline we use a phrase-based SMT model. The MT models are fine-tuned on the development set, and the best configurations are used to report the evaluation on the test set. Experiments show the BLEU of 13.71 on the benchmark test data. This is better compared to the official baseline BLEU score of 10.79.
To enhance Neural Machine Translation models, several obvious ways such as enlarging the hidden size of recurrent layers and stacking multiple layers of RNN can be considered. Surprisingly, we observe that using naively stacked RNNs in the decoder slows down the training and leads to degradation in performance. In this paper, We demonstrate that applying residual connections in the depth of stacked RNNs can help the optimization, which is referred to as residual stacking. In empirical evaluation, residual stacking of decoder RNNs gives superior results compared to other methods of enhancing the model with a fixed parameter budget. Our submitted systems in WAT2016 are based on a NMT model ensemble with residual stacking in the decoder. To further improve the performance, we also attempt various methods of system combination in our experiments.
Cancer (a.k.a neoplasms in a broader sense) is one of the leading causes of death worldwide and its incidence is expected to exacerbate. To respond to the critical need from the society, there have been rigorous attempts for the cancer research community to develop treatment for cancer. Accordingly, we observe a surge in the sheer volume of research products and outcomes in relation to neoplasms. In this talk, we introduce the notion of entitymetrics to provide a new lens for understanding the impact, trend, and diffusion of knowledge associated with neoplasms research. To this end, we collected over two million records from PubMed, the most popular search engine in the medical domain. Coupled with text mining techniques including named entity recognition, sentence boundary detection, string approximate matching, entitymetrics enables us to analyze knowledge diffusion, impact, and trend at various knowledge entity units, such as bio-entity, organization, and country. At the end of the talk, the future applications and possible directions of entitymetrics will be discussed.
The present paper explores a novel method that integrates efficient distributed representations with terminology extraction. We show that the information from a small number of observed instances can be combined with local and global word embeddings to remarkably improve the term extraction results on unigram terms. To do so we pass the terms extracted by other tools to a filter made of the local-global embeddings and a classifier which in turn decides whether or not a term candidate is a term. The filter can also be used as a hub to merge different term extraction tools into a single higher-performing system. We compare filters that use the skip-gram architecture and filters that employ the CBOW architecture for the task at hand.
In the paper, we address the problem of recognition of non-domain phrases in terminology lists obtained with an automatic term extraction tool. We focus on identification of multi-word phrases that are general terms and discourse function expressions. We tested several methods based on domain corpora comparison and a method based on contexts of phrases identified in a large corpus of general language. We compared the results of the methods to manual annotation. The results show that the task is quite hard as the inter-annotator agreement is low. Several tested methods achieved similar overall results, although the phrase ordering varied between methods. The most successful method with the precision about 0.75 at the half of the tested list was the context based method using a modified contextual diversity coefficient.
This article presents a domain-driven algorithm for the task of term sense disambiguation (TSD). TSD aims at automatically choosing which term record from a term bank best represents the meaning of a term occurring in a particular context. In a translation environment, finding the contextually appropriate term record is necessary to access the proper equivalent to be used in the target language text. The term bank TERMIUM Plus, recently published as an open access repository, is chosen as a domain-rich resource for testing our TSD algorithm, using English and French as source and target languages. We devise an experiment using over 1300 English terms found in scientific articles, and show that our domain-driven TSD algorithm is able to bring the best term record, and therefore the best French equivalent, at the average rank of 1.69 compared to a baseline random rank of 3.51.
In this paper, we propose a method of augmenting existing bilingual terminologies. Our method belongs to a “generate and validate” framework rather than extraction from corpora. Although many studies have proposed methods to find term translations or to augment terminology within a “generate and validate” framework, few has taken full advantage of the systematic nature of terminologies. A terminology of a domain represents the conceptual system of the domain fairly systematically, and we contend that making use of the systematicity fully will greatly contribute to the effective augmentation of terminologies. This paper proposes and evaluates a novel method to generate bilingual term candidates by using existing terminologies and delving into their systematicity. Experiments have shown that our method can generate much better term candidate pairs than the existing method and give improved performance for terminology augmentation.
The extraction of data exemplifying relations between terms can make use, at least to a large extent, of techniques that are similar to those used in standard hybrid term candidate extraction, namely basic corpus analysis tools (e.g. tagging, lemmatization, parsing), as well as morphological analysis of complex words (compounds and derived items). In this article, we discuss the use of such techniques for the extraction of raw material for a description of relations between terms, and we provide internal evaluation data for the devices developed. We claim that user-generated content is a rich source of term variation through paraphrasing and reformulation, and that these provide relational data at the same time as term variants. Germanic languages with their rich word formation morphology may be particularly good candidates for the approach advocated here.
We investigate how both model-related factors and application-related factors affect the accuracy of distributional semantic models (DSMs) in the context of specialized lexicography, and how these factors interact. This holistic approach to the evaluation of DSMs provides valuable guidelines for the use of these models and insight into the kind of semantic information they capture.
We propose and evaluate a method for identifying co-hyponym lexical units in a terminological resource. The principles of term recognition and distributional semantics are combined to extract terms from a similar category of concept. Given a set of candidate terms, random projections are employed to represent them as low-dimensional vectors. These vectors are derived automatically from the frequency of the co-occurrences of the candidate terms and words that appear within windows of text in their proximity (context-windows). In a k-nearest neighbours framework, these vectors are classified using a small set of manually annotated terms which exemplify concept categories. We then investigate the interplay between the size of the corpus that is used for collecting the co-occurrences and a number of factors that play roles in the performance of the proposed method: the configuration of context-windows for collecting co-occurrences, the selection of neighbourhood size (k), and the choice of similarity metric.
Despite advances in computer technology, terminologists still tend to rely on manual work to extract all the semantic information that they need for the description of specialized concepts. In this paper we propose the creation of new word sketches in Sketch Engine for the extraction of semantic relations. Following a pattern-based approach, new sketch grammars are devel-oped in order to extract some of the most common semantic relations used in the field of ter-minology: generic-specific, part-whole, location, cause and function.
This paper presents the construction and evaluation of Japanese and English controlled bilingual terminologies that are particularly intended for controlled authoring and machine translation with special reference to the Japanese municipal domain. Our terminologies are constructed by extracting terms from municipal website texts, and the term variations are controlled by defining preferred and proscribed terms for both the source Japanese and the target English. To assess the coverage of the terms/concepts in the municipal domain and validate the quality of the control, we employ a quantitative extrapolation method that estimates the potential vocabulary size. Using Large-Number-of-Rare-Event (LNRE) modelling, we compare two parameters: (1) uncontrolled and controlled and (2) Japanese and English. The results show that our terminologies currently cover about 45–65% of the terms and 50–65% of the concepts in the municipal domain, and are well controlled. The detailed analysis of growth patterns of terminologies also provides insight into the extent to which we can enlarge the terminologies within the realistic range.
By its own nature, the Natural Language Processing (NLP) community is a priori the best equipped to study the evolution of its own publications, but works in this direction are rare and only recently have we seen a few attempts at charting the field. In this paper, we use the algorithms, resources, standards, tools and common practices of the NLP field to build a list of terms characteristic of ongoing research, by mining a large corpus of scientific publications, aiming at the largest possible exhaustivity and covering the largest possible time span. Study of the evolution of this term list through time reveals interesting insights on the dynamics of field and the availability of the term database and of the corpus (for a large part) make possible many further comparative studies in addition to providing a test field for a new graphic interface designed to perform visual time analytics of large sized thesauri.
Annotating medical text such as clinical notes with human phenotype descriptors is an important task that can, for example, assist in building patient profiles. To automatically annotate text one usually needs a dictionary of predefined terms. However, do to the variety of human expressiveness, current state-of-the art phenotype concept recognizers and automatic annotators struggle with specific domain issues and challenges. In this paper we present results of an-notating gold standard corpus with a dictionary containing lexical variants for the Human Phenotype Ontology terms. The main purpose of the dictionary is to improve the recall of phenotype concept recognition systems. We compare the method with four other approaches and present results.
We propose a semi-automatic method for the acquisition of specialised ontological and terminological knowledge. An ontology and a terminology are automatically built from domain experts’ annotations. The ontology formalizes the common and shared conceptual vocabulary of those experts. Its associated terminology defines a glossary linking annotated terms to their semantic categories. These two resources evolve incrementally and are used for an automatic annotation of a new corpus at each iteration. The annotated corpus concerns the evaluation of French higher education and science institutions.
With many hospitals digitalizing clinical records it has opened opportunities for researchers in NLP, Machine Learning to apply techniques for extracting meaning and make actionable insights. There has been previous attempts in mapping free text to medical nomenclature like UMLS, SNOMED. However, in this paper, we had analyzed diagnosis in clinical reports using ICD10 to achieve a lightweight, real-time predictions by introducing concepts like WordInfo, root word identification. We were able to achieve 68.3% accuracy over clinical records collected from qualified clinicians. Our study would further help the healthcare institutes in organizing their clinical reports based on ICD10 mappings and derive numerous insights to achieve operational efficiency and better medical care.