This work compares the performances achieved by Phrase-Based Statistical Machine Translation systems (PB-SMT) and attention-based Neuronal Machine Translation systems (NMT) when translating User Generated Content (UGC), as encountered in social medias, from French to English. We show that, contrary to what could be expected, PBSMT outperforms NMT when translating non-canonical inputs. Our error analysis uncovers the specificities of UGC that are problematic for sequential NMT architectures and suggests new avenue for improving NMT models.

Standard approaches to treebanking traditionally employ a waterfall model (Sommerville, 2010), where annotation guidelines guide the annotation process and insights from the annotation process in turn lead to subsequent changes in the annotation guidelines. This process remains a very expensive step in creating linguistic resources for a target language, necessitates both linguistic expertise and manual effort to develop the annotations and is subject to inconsistencies in the annotation due to human errors. In this paper, we propose an alternative approach to treebanking—one that requires writing grammars. This approach is motivated specifically in the context of Universal Dependencies, an effort to develop uniform and cross-lingually consistent treebanks across multiple languages. We show here that a bootstrapping approach to treebanking via interlingual grammars is plausible and useful in a process where grammar engineering and treebanking are jointly pursued when creating resources for the target language. We demonstrate the usefulness of synthetic treebanks in the task of delexicalized parsing. Our experiments reveal that simple models for treebank generation are cheaper than human annotated treebanks, especially in the lower ends of the learning curves for delexicalized parsing, which is relevant in particular in the context of low-resource languages.

More and more evidence is appearing that integrating symbolic lexical knowledge into neural models aids learning. This contrasts the widely-held belief that neural networks largely learn their own feature representations. For example, recent work has shows benefits of integrating lexicons to aid cross-lingual part-of-speech (PoS). However, little is known on how complementary such additional information is, and to what extent improvements depend on the coverage and quality of these external resources. This paper seeks to fill this gap by providing a thorough analysis on the contributions of lexical resources for cross-lingual PoS tagging in neural times.

This paper investigates the presence of gender bias in pretrained Swedish embeddings. We focus on a scenario where names are matched with occupations, and we demonstrate how a number of standard pretrained embeddings handle this task. Our experiments show some significant differences between the pretrained embeddings, with word-based methods showing the most bias and contextualized language models showing the least. We also demonstrate that the previously proposed debiasing method does not affect the performance of the various embeddings in this scenario.

Determining how words have changed their meaning is an important topic in Natural Language Processing. However, evaluations of methods to characterise such change have been limited to small, handcrafted resources. We introduce an English evaluation set which is larger, more varied, and more realistic than seen to date, with terms derived from a historical thesaurus. Moreover, the dataset is unique in that it represents change as a shift from the term of interest to a WordNet synset. Using the synset lemmas, we can use this set to evaluate (standard) methods that detect change between word pairs, as well as (adapted) methods that detect the change between a term and a sense overall. We show that performance on the new data set is much lower than earlier reported findings, setting a new standard.

We apply hyperbolic embeddings to trace the dynamics of change of conceptual-semantic relationships in a large diachronic scientific corpus (200 years). Our focus is on emerging scientific fields and the increasingly specialized terminology establishing around them. Reproducing high-quality hierarchical structures such as WordNet on a diachronic scale is a very difficult task. Hyperbolic embeddings can map partial graphs into low dimensional, continuous hierarchical spaces, making more explicit the latent structure of the input. We show that starting from simple lists of word pairs (rather than a list of entities with directional links) it is possible to build diachronic hierarchical semantic spaces which allow us to model a process towards specialization for selected scientific fields.

We present an evaluation of Czech low-dimensional distributed word representations, also known as word embeddings. We describe five different approaches to training the models and three different corpora used in training. We evaluate the resulting models on five different datasets, report the results and provide their further analysis.

In this paper, we investigate the effect of enhancing lexical embeddings in LSTM language models (LM) with syntactic and semantic representations. We evaluate the language models using perplexity, and we evaluate the performance of the models on the task of predicting human sentence acceptability judgments. We train LSTM language models on sentences automatically annotated with universal syntactic dependency roles (Nivre, 2016), dependency depth and universal semantic tags (Abzianidze et al., 2017) to predict sentence acceptability judgments. Our experiments indicate that syntactic tags lower perplexity, while semantic tags increase it. Our experiments also show that neither syntactic nor semantic tags improve the performance of LSTM language models on the task of predicting sentence acceptability judgments.

In this paper, we compare the use of linear versus neural classifiers in a greedy transition system for MWE identification. Both our linear and neural models achieve a new state-of-the-art on the PARSEME 1.1 shared task data sets, comprising 20 languages. Surprisingly, our best model is a simple feed-forward network with one hidden layer, although more sophisticated (recurrent) architectures were tested. The feedback from this study is that tuning a SVM is rather straightforward, whereas tuning our neural system revealed more challenging. Given the number of languages and the variety of linguistic phenomena to handle for the MWE identification task, we have designed an accurate tuning procedure, and we show that hyperparameters are better selected by using a majority-vote within random search configurations rather than a simple best configuration selection. Although the performance is rather good (better than both the best shared task system and the average of the best per-language results), further work is needed to improve the generalization power, especially on unseen MWEs.

This paper analyzes results on light-verb construction identification from the PARSEME shared-task, distinguishing between simple cases that could be directly learned from training data from more complex cases that require an extra level of semantic processing. We propose a simple baseline that beats the state of the art for the simple cases, and couple it with another simple baseline to handle the complex cases. We additionally present two other classifiers based on a richer set of features, with results surpassing the state of the art by 8 percentage points.

We explore the effectiveness of four feature representations – bag-of-words, word embeddings, principal components and autoencoders – for the binary categorization of the easy-to-read variety vs standard language. Standard language refers to the ordinary language variety used by a population as a whole or by a community, while the “easy-to-read” variety is a simpler (or a simplified) version of the standard language. We test the efficiency of these feature representations on three corpora, which differ in size, class balance, unit of analysis, language and topic. We rely on supervised and unsupervised machine learning algorithms. Results show that bag-of-words is a robust and straightforward feature representation for this task and performs well in many experimental settings. Its performance is equivalent or equal to the performance achieved with principal components and autoencorders, whose preprocessing is however more time-consuming. Word embeddings are less accurate than the other feature representations for this classification task.

Affordances denote actions that can be performed in the presence of different objects, or possibility of action in an environment. In robotic systems, affordances and actions may suffer from poor semantic generalization capabilities due to the high amount of required hand-crafted specifications. To alleviate this issue, we propose a method to mine for object-action pairs in free text corpora, successively training and evaluating different prediction models of affordance based on word embeddings.

This paper documents the creation of a large-scale dataset of evaluative sentences – i.e. both subjective and objective sentences that are found to be sentiment-bearing – based on mixed-domain professional reviews from various news-sources. We present both the annotation scheme and first results for classification experiments. The effort represents a step toward creating a Norwegian dataset for fine-grained sentiment analysis.

We present our work towards developing a system that should find, in a large text corpus, contiguous phrases expressing similar meaning as a query phrase of arbitrary length. Depending on the use case, this task can be seen as a form of (phrase-level) query rewriting. The suggested approach works in a generative manner, is unsupervised and uses a combination of a semantic word n-gram model, a statistical language model and a document search engine. A central component is a distributional semantic model containing word n-grams vectors (or embeddings) which models semantic similarities between n-grams of different order. As data we use a large corpus of PubMed abstracts. The presented experiment is based on manual evaluation of extracted phrases for arbitrary queries provided by a group of evaluators. The results indicate that the proposed approach is promising and that the use of distributional semantic models trained with uni-, bi- and trigrams seems to work better than a more traditional unigram model.

We present ParIce, a new English-Icelandic parallel corpus. This is the first parallel corpus built for the purposes of language technology development and research for Icelandic, although some Icelandic texts can be found in various other multilingual parallel corpora. We map out which Icelandic texts are available for these purposes, collect aligned data and align other bilingual texts we acquired. We describe the alignment process and how we filter the data to weed out noise and bad alignments. In total we collected 43 million Icelandic words in 4.3 million aligned segment pairs, but after filtering, our corpus includes 38.8 million Icelandic words in 3.5 million segment pairs. We estimate that approximately 5% of the corpus data is noise or faulty alignments while more than 50% of the segments we deleted were faulty. We estimate that our filtering process reduced the number of faulty segments in the corpus by more than 60% while only reducing the number of good alignments by approximately 8%.

The topic of this paper is The Database of Icelandic Morphology (DIM), a multipurpose linguistic resource, created for use in language technology, as a reference for the general public in Iceland, and for use in research on the Icelandic language. DIM contains inflectional paradigms and analysis of word formation, with a vocabulary of approx. 285,000 lemmas. DIM is based on The Database of Modern Icelandic Inflection, which has been in use since 2004.

This paper presents an overview of the available linguistic resources for the Sakha language, and presents new tools for supporting language learning for Sakha. The essential resources include a morphological analyzer, digital dictionaries, and corpora of Sakha texts. Based on these resources, we implement a language-learning environment for Sakha in the Revita CALL platform. We extended an earlier, preliminary version of the morphological analyzer/transducer, built on the Apertium finite-state platform. The analyzer currently has an adequate level of coverage, between 86% and 89% on two Sakha corpora. Revita is a freely available online language learning platform for learners beyond the beginner level. We describe the tools for Sakha currently integrated into the Revita platform. To the best of our knowledge, at present, this is the first large-scale project undertaken to support intermediate-advanced learners of a minority Siberian language.

We show how to express the problem of finding an optimal morpheme segmentation from a set of labelled words as a 0/1 linear programming problem, and how to build on this to analyse a language’s morphology. The approach works even when there is very little training data available.

This paper explores the use of multi-task learning (MTL) for incorporating external knowledge in neural models. Specifically, we show how MTL can enable a BiLSTM sentiment classifier to incorporate information from sentiment lexicons. Our MTL set-up is shown to improve model performance (compared to a single-task set-up) on both English and Norwegian sentence-level sentiment datasets. The paper also introduces a new sentiment lexicon for Norwegian.

Sentiment analysis has become very popular in both research and business due to the increasing amount of opinionated text from Internet users. Standard sentiment analysis deals with classifying the overall sentiment of a text, but this doesn’t include other important information such as towards which entity, topic or aspect within the text the sentiment is directed. Aspect-based sentiment analysis (ABSA) is a more complex task that consists in identifying both sentiments and aspects. This paper shows the potential of using the contextual word representations from the pre-trained language model BERT, together with a fine-tuning method with additional generated text, in order to solve out-of-domain ABSA and outperform previous state-of-the-art results on SemEval-2015 Task 12 subtask 2 and SemEval-2016 Task 5. To the best of our knowledge, no other existing work has been done on out-of-domain ABSA for aspect classification.

The task of stance detection consists of classifying the opinion within a text towards some target. This paper seeks to generate a dataset of quotes from Danish politicians, label this dataset to allow the task of stance detection to be performed, and present annotation guidelines to allow further expansion of the generated dataset. Furthermore, three models based on an LSTM architecture are designed, implemented and optimized to perform the task of stance detection for the generated dataset. Experiments are performed using conditionality and bi-directionality for these models, and using either singular word embeddings or averaged word embeddings for an entire quote, to determine the optimal model design. The simplest model design, applying neither conditionality or bi-directionality, and averaged word embeddings across quotes, yields the strongest results. Furthermore, it was found that inclusion of the quotes politician, and the party affiliation of the quoted politician, greatly improved performance of the strongest model.

The net is rife with rumours that spread through microblogs and social media. Not all the claims in these can be verified. However, recent work has shown that the stances alone that commenters take toward claims can be sufficiently good indicators of claim veracity, using e.g. an HMM that takes conversational stance sequences as the only input. Existing results are monolingual (English) and mono-platform (Twitter). This paper introduces a stance-annotated Reddit dataset for the Danish language, and describes various implementations of stance classification models. Of these, a Linear SVM provides predicts stance best, with 0.76 accuracy / 0.42 macro F1. Stance labels are then used to predict veracity across platforms and also across languages, training on conversations held in one language and using the model on conversations held in another. In our experiments, monolinugal scores reach stance-based veracity accuracy of 0.83 (F1 0.68); applying the model across languages predicts veracity of claims with an accuracy of 0.82 (F1 0.67). This demonstrates the surprising and powerful viability of transferring stance-based veracity prediction across languages.

NER is the task of recognizing and demarcating the segments of a document that are part of a name and which type of name it is. We use 4 different categories of names: Locations (LOC), miscellaneous (MISC), organizations (ORG), and persons (PER). Even though we employ state of the art methods—including sub-word embeddings—that work well for English, we are unable to reproduce the same success for the Norwegian written forms. However, our model performs better than any previous research on Norwegian text. The study also presents the first NER for Nynorsk. Lastly, we find that by combining Nynorsk and Bokmål into one training corpus we improve the performance of our model on both languages.

Named entity recognition (NER) is a well-researched task in the field of NLP, which typically requires large annotated corpora for training usable models. This is a problem for languages which lack large annotated corpora, such as Finnish. We propose an approach to create a named entity recognizer with no annotated or parallel documents, by leveraging strong NER models that exist for English. We automatically gather a large amount of chronologically matched data in two languages, then project named entity annotations from the English documents onto the Finnish ones, by resolving the matches with limited linguistic rules. We use this “artificially” annotated data to train a BiLSTM-CRF model. Our results show that this method can produce annotated instances with high precision, and the resulting model achieves state-of-the-art performance.

News articles such as sports game reports are often thought to closely follow the underlying game statistics, but in practice they contain a notable amount of background knowledge, interpretation, insight into the game, and quotes that are not present in the official statistics. This poses a challenge for automated data-to-text news generation with real-world news corpora as training data. We report on the development of a corpus of Finnish ice hockey news, edited to be suitable for training of end-to-end news generation methods, as well as demonstrate generation of text, which was judged by journalists to be relatively close to a viable product. The new dataset and system source code are available for research purposes.

Historical cryptology is the study of historical encrypted messages aiming at their decryption by analyzing the mathematical, linguistic and other coding patterns and their historical context. In libraries and archives we can find quite a lot of ciphers, as well as keys describing the method used to transform the plaintext message into a ciphertext. In this paper, we present work on automatically mapping keys to ciphers to reconstruct the original plaintext message, and use language models generated from historical texts to guess the underlying plaintext language.

Human voice provides the means for verbal communication and forms a part of personal identity. Due to genetic and environmental factors, a voice of a child should resemble the voice of her parent(s), but voice similarities between parents and young children are underresearched. Read-aloud speech of Finnish-speaking and Russian-speaking parent-child pairs was subject to perceptual and multi-step instrumental and statistical analysis. Finnish-speaking listeners could not discriminate family pairs auditorily in an XAB paradigm, but the Russian-speaking listeners’ mean accuracy of answers reached 72.5%. On average, in both language groups family-internal f0 similarities were stronger than family-external, with parents showing greater family-internal similarities than children. Auditory similarities did not reflect acoustic similarities in a straightforward way.

Cross-modality between vision and language is a key component for effective and efficient communication, and human language processing mechanism successfully integrates information from various modalities to extract the intended meaning. However, incomplete linguistic input, i.e. due to a noisy environment, is one of the challenges for a successful communication. In that case, an incompleteness in one channel can be compensated by information from another one. In this paper, by conducting visual-world paradigm, we investigated the dynamics between syntactically possible gap fillers and the visual arrangements in incomplete German sentences and their effect on overall sentence interpretation.

In this paper we introduce a new natural language processing dataset and benchmark for predicting prosodic prominence from written text. To our knowledge this will be the largest publicly available dataset with prosodic labels. We describe the dataset construction and the resulting benchmark dataset in detail and train a number of different models ranging from feature-based classifiers to neural network systems for the prediction of discretized prosodic prominence. We show that pre-trained contextualized word representations from BERT outperform the other models even with less than 10% of the training data. Finally we discuss the dataset in light of the results and point to future research and plans for further improving both the dataset and methods of predicting prosodic prominence from text. The dataset and the code for the models will be made publicly available.

We consider cross- and multilingual text classification approaches to the identification of online registers (genres), i.e. text varieties with specific situational characteristics. Register is the most important predictor of linguistic variation, and register information could improve the potential of online data for many applications. We introduce the first manually annotated non-English corpus of online registers featuring the full range of linguistic variation found online. The data set consists of 2,237 Finnish documents and follows the register taxonomy developed for the Corpus of Online Registers of English (CORE). Using CORE and the newly introduced corpus, we demonstrate the feasibility of cross-lingual register identification using a simple approach based on convolutional neural networks and multilingual word embeddings. We further find that register identification results can be improved through multilingual training even when a substantial number of annotations is available in the target language.

We present a system for Natural Language Inference which uses a dynamic semantics converter from abstract syntax trees to Coq types. It combines the fine-grainedness of a dynamic semantics system with the powerfulness of a state-of-the-art proof assistant, like Coq. We evaluate the system on all sections of the FraCaS test suite, excluding section 6. This is the first system that does a complete run on the anaphora and ellipsis sections of the FraCaS. It has a better overall accuracy than any previous system.

We investigate different ensemble learning techniques for neural morphological inflection using bidirectional LSTM encoder-decoder models with attention. We experiment with weighted and unweighted majority voting and bagging. We find that all investigated ensemble methods lead to improved accuracy over a baseline of a single model. However, contrary to expectation based on earlier work by Najafi et al. (2018) and Silfverberg et al. (2017), weighting does not deliver clear benefits. Bagging was found to underperform plain voting ensembles in general.

Lemmatization, finding the basic morphological form of a word in a corpus, is an important step in many natural language processing tasks when working with morphologically rich languages. We describe and evaluate Nefnir, a new open source lemmatizer for Icelandic. Nefnir uses suffix substitution rules, derived from a large morphological database, to lemmatize tagged text. Evaluation shows that for correctly tagged text, Nefnir obtains an accuracy of 99.55%, and for text tagged with a PoS tagger, the accuracy obtained is 96.88%.

Social science researchers often use text as the raw data in investigations: for instance, when investigating the effects of IMF policies on the development of countries under IMF programs, researchers typically encode structured descriptions of the programs using a time-consuming manual effort. Making this process automatic may open up new opportunities in scaling up such investigations. As a first step towards automatizing this coding process, we describe an experiment where we apply a sentence classifier that automatically detects mentions of policy conditions in IMF loan agreements and divides them into different types. The results show that the classifier is generally able to detect the policy conditions, although some types are hard to distinguish.

In this paper, we present a prototype for an online exercise aimed at learners of English and Swedish that serves multiple purposes. The exercise allows learners of the aforementioned languages to train their knowledge of particle verbs receiving clues from the exercise application. The user themselves decide which clue to receive and pay in virtual currency for each, which provides us with valuable information about the utility of the clues that we provide as well as the learners willingness to trade virtual currency versus accuracy of their choice. As resources, we use list with annotated levels from the proficiency scale defined by the Common European Framework of Reference (CEFR) and a multilingual corpus with syntactic dependency relations and word annotation for all language pairs. From the latter resource, we extract translation equivalents for particle verb construction together with a list of parallel corpus examples that can be used as clues in the exercise.

In this paper we present a new method to learn a model robust to typos for a Named Entity Recognition task. Our improvement over existing methods helps the model to take into account the context of the sentence inside a justice decision in order to recognize an entity with a typo. We used state-of-the-art models and enriched the last layer of the neural network with high-level information linked with the potential of the word to be a certain type of entity. More precisely, we utilized the similarities between the word and the potential entity candidates the tagged sentence context. The experiments on a dataset of french justice decisions show a reduction of the relative F1-score error of 32%, upgrading the score obtained with the most competitive fine-tuned state-of-the-art system from 94.85% to 96.52%.

In this paper, we present a Bayesian approach to natural language semantics. Our main focus is on the inference task in an environment where judgments require probabilistic reasoning. We treat nouns, verbs, adjectives, etc. as unary predicates, and we model them as boxes in a bounded domain. We apply Bayesian learning to satisfy constraints expressed as premises. In this way we construct a model, by specifying boxes for the predicates. The probability of the hypothesis (the conclusion) is evaluated against the model that incorporates the premises as constraints.

This paper introduces language processing resources and tools for Bornholmsk, a language spoken on the island of Bornholm, with roots in Danish and closely related to Scanian. This presents an overview of the language and available data, and the first NLP models for this living, minority Nordic language. Sammenfattnijng på borrijnholmst: Dæjnna artikkelijn introduserer natursprågsresurser å varktoi for borrijnholmst, ed språg a dær snakkes på ön Borrijnholm me rødder i danst å i nær familia me skånst. Artikkelijn gjer ed âuersyn âuer språged å di datan som fijnnes, å di fosste NLP modællarna for dætta læwenes nordiska minnretâlsspråged.

Endangered Uralic languages present a high variety of inflectional forms in their morphology. This results in a high number of homonyms in inflections, which introduces a lot of morphological ambiguity in sentences. Previous research has employed constraint grammars to address this problem, however CGs are often unable to fully disambiguate a sentence, and their development is labour intensive. We present an LSTM based model for automatically ranking morphological readings of sentences based on their quality. This ranking can be used to evaluate the existing CG disambiguators or to directly morphologically disambiguate sentences. Our approach works on a morphological abstraction and it can be trained with a very small dataset.

This paper describes an evaluation of five data-driven part-of-speech (PoS) taggers for spoken Norwegian. The taggers all rely on different machine learning mechanisms: decision trees, hidden Markov models (HMMs), conditional random fields (CRFs), long-short term memory networks (LSTMs), and convolutional neural networks (CNNs). We go into some of the challenges posed by the task of tagging spoken, as opposed to written, language, and in particular a wide range of dialects as is found in the recordings of the LIA (Language Infrastructure made Accessible) project. The results show that the taggers based on either conditional random fields or neural networks perform much better than the rest, with the LSTM tagger getting the highest score.

Danish is a North Germanic language spoken principally in Denmark, a country with a long tradition of technological and scientific innovation. However, the language has received relatively little attention from a technological perspective. In this paper, we review Natural Language Processing (NLP) research, digital resources and tools which have been developed for Danish. We find that availability of models and tools is limited, which calls for work that lifts Danish NLP a step closer to the privileged languages. Dansk abstrakt: Dansk er et nordgermansk sprog, talt primært i kongeriget Danmark, et land med stærk tradition for teknologisk og videnskabelig innovation. Det danske sprog har imidlertid været genstand for relativt begrænset opmærksomhed, teknologisk set. I denne artikel gennemgår vi sprogteknologi-forskning, -ressourcer og -værktøjer udviklet for dansk. Vi konkluderer at der eksisterer et fåtal af modeller og værktøjer, hvilket indbyder til forskning som løfter dansk sprogteknologi i niveau med mere priviligerede sprog.

We report on work in progress which consists of annotating an Icelandic corpus for named entities (NEs) and using it for training a named entity recognizer based on a Bidirectional Long Short-Term Memory model. Currently, we have annotated 7,538 NEs appearing in the first 200,000 tokens of a 1 million token corpus, MIM-GOLD, originally developed for serving as a gold standard for part-of-speech tagging. Our best performing model, trained on this subset of MIM-GOLD, and enriched with external word embeddings, obtains an overall F1 score of 81.3% when categorizing NEs into the following four categories: persons, locations, organizations and miscellaneous. Our preliminary results are promising, especially given the fact that 80% of MIM-GOLD has not yet been used for training.

Named Entity Recognition (NER) has greatly advanced by the introduction of deep neural architectures. However, the success of these methods depends on large amounts of training data. The scarcity of publicly-available human-labeled datasets has resulted in limited evaluation of existing NER systems, as is the case for Danish. This paper studies the effectiveness of cross-lingual transfer for Danish, evaluates its complementarity to limited gold data, and sheds light on performance of Danish NER.

The use of a linking element between compound members is a common phenomenon in Germanic languages. Still, the exact use and conditioning of such elements is a disputed topic in linguistics. In this paper we address the issue of predicting the use of linking elements in Danish. Following previous research that shows how the choice of linking element might be conditioned by phonology, we frame the problem as a language modeling task: Considering the linking elements -s/-∅ the problem becomes predicting what is most probable to encounter next, a syllable boundary or the joining element, ‘s’. We show that training a language model on this task reaches an accuracy of 94 %, and in the case of an unsupervised model, the accuracy reaches 80%.

This article is a report from an ongoing project aiming at analyzing lexical and grammatical competences of Swedish as a Second language (L2). To facilitate lexical analysis, we need access to metalinguistic information about relevant vocabulary that L2 learners can use and understand. The focus of the current article is on the lexical annotation of the vocabulary scope for a range of lexicographical aspects, such as morphological analysis, valency, types of multi-word units, etc. We perform parts of the analysis automatically, and other parts manually. The rationale behind this is that where there is no possibility to add information automatically, manual effort needs to be added. To facilitate the latter, a tool LEGATO has been designed, implemented and currently put to active testing.

This paper presents a flexible and powerful system for creating parallel corpora and for running neural machine translation services. Our package provides a scalable data repository backend that offers transparent data pre-processing pipelines and automatic alignment procedures that facilitate the compilation of extensive parallel data sets from a variety of sources. Moreover, we develop a web-based interface that constitutes an intuitive frontend for end-users of the platform. The whole system can easily be distributed over virtual machines and implements a sophisticated permission system with secure connections and a flexible database for storing arbitrary metadata. Furthermore, we also provide an interface for neural machine translation that can run as a service on virtual machines, which also incorporates a connection to the data repository software.

We present a new method for preparing a lexical-phonetic database as a resource for acoustic model training. The research is an offshoot of the ongoing Project Ravnur (Speech Recognition for Faroese), but the method is language-independent. At NODALIDA 2019 we demonstrate the method (called SHARP) online, showing how a traditional lexical-phonetic dictionary (with a very rich phone inventory) is transformed into an ASR-friendly database (with reduced phonetics, preventing data sparseness). The mapping procedure is informed by a corpus of speech transcripts. We conclude with a discussion on the benefits of a well-thought-out BLARK design (Basic Language Resource Kit), making tools like SHARP possible.

The availability of user-generated content has increased significantly over time. Wikipedia is one example of a corpora which spans a huge range of topics and is freely available. Storing and processing these corpora requires flexible documents models as they may contain malicious and incorrect data. Docria is a library which attempts to address this issue by providing a solution which can be used with small to large corpora, from laptops using Python interactively in a Jupyter notebook to clusters running map-reduce frameworks with optimized compiled code. Docria is available as open-source code.

This paper describes the design and use of the graph-based parsing framework and toolkit UniParse, released as an open-source python software package. UniParse as a framework novelly streamlines research prototyping, development and evaluation of graph-based dependency parsing architectures. UniParse does this by enabling highly efficient, sufficiently independent, easily readable, and easily extensible implementations for all dependency parser components. We distribute the toolkit with ready-made configurations as re-implementations of all current state-of-the-art first-order graph-based parsers, including even more efficient Cython implementations of both encoders and decoders, as well as the required specialised loss functions.

Defining words in a textual context is a useful task both for practical purposes and for gaining insight into distributed word representations. Building on the distributional hypothesis, we argue here that the most natural formalization of definition modeling is to treat it as a sequence-to-sequence task, rather than a word-to-sequence task: given an input sequence with a highlighted word, generate a contextually appropriate definition for it. We implement this approach in a Transformer-based sequence-to-sequence model. Our proposal allows to train contextualization and definition generation in an end-to-end fashion, which is a conceptual improvement over earlier works. We achieve state-of-the-art results both in contextual and non-contextual definition modeling.

We extend a state-of-the-art deep neural architecture for semantic dependency parsing with features defined over syntactic dependency trees. Our empirical results show that only gold-standard syntactic information leads to consistent improvements in semantic parsing accuracy, and that the magnitude of these improvements varies with the specific combination of the syntactic and the semantic representation used. In contrast, automatically predicted syntax does not seem to help semantic parsing. Our error analysis suggests that there is a significant overlap between syntactic and semantic representations.

In this paper, we critically evaluate the widespread assumption that deep learning NLP models do not require lemmatized input. To test this, we trained versions of contextualised word embedding ELMo models on raw tokenized corpora and on the corpora with word tokens replaced by their lemmas. Then, these models were evaluated on the word sense disambiguation task. This was done for the English and Russian languages. The experiments showed that while lemmatization is indeed not necessary for English, the situation is different for Russian. It seems that for rich-morphology languages, using lemmatized training and testing data yields small but consistent improvements: at least for word sense disambiguation. This means that the decisions about text pre-processing before training ELMo should consider the linguistic nature of the language in question.

The multilingual BERT model is trained on 104 languages and meant to serve as a universal language model and tool for encoding sentences. We explore how well the model performs on several languages across several tasks: a diagnostic classification probing the embeddings for a particular syntactic property, a cloze task testing the language modelling ability to fill in gaps in a sentence, and a natural language generation task testing for the ability to produce coherent text fitting a given context. We find that the currently available multilingual BERT model is clearly inferior to the monolingual counterparts, and cannot in many cases serve as a substitute for a well-trained monolingual model. We find that the English and German models perform well at generation, whereas the multilingual model is lacking, in particular, for Nordic languages. The code of the experiments in the paper is available at: https://github.com/TurkuNLP/bert-eval

Encoders that generate representations based on context have, in recent years, benefited from adaptations that allow for pre-training on large text corpora. Earlier work on evaluating fixed-length sentence representations has included the use of ‘probing’ tasks, that use diagnostic classifiers to attempt to quantify the extent to which these encoders capture specific linguistic phenomena. The principle of probing has also resulted in extended evaluations that include relatively newer word-level pre-trained encoders. We build on probing tasks established in the literature and comprehensively evaluate and analyse – from a typological perspective amongst others – multilingual variants of existing encoders on probing datasets constructed for 6 non-English languages. Specifically, we probe each layer of a multiple monolingual RNN-based ELMo models, the transformer-based BERT’s cased and uncased multilingual variants, and a variant of BERT that uses a cross-lingual modelling scheme (XLM).

Due to the differences between reviews in different product categories, creating a general model for cross-domain sentiment classification can be a difficult task. This paper proposes an architecture that incorporates domain knowledge into a neural sentiment classification model. In addition to providing a cross-domain model, this also provides a quantifiable representation of the domains as numeric vectors. We show that it is possible to cluster the domain vectors and provide qualitative insights into the inter-domain relations. We also a) present a new data set for sentiment classification that includes a domain parameter and preprocessed data points, and b) perform an ablation study in order to determine whether some word groups impact performance.

This paper discusses methods to improve the performance of text classification on data that is difficult to classify due to a large number of unbalanced classes with noisy examples. A variety of features are tested, in combination with three different neural-network-based methods with increasing complexity. The classifiers are applied to a songtext–artist dataset which is large, unbalanced and noisy. We come to the conclusion that substantial improvement can be obtained by removing unbalancedness and sparsity from the data. This fulfils a classification task unsatisfactorily—however, with contemporary methods, it is a practical step towards fairly satisfactory results.