Corpora are important assets within the natural language processing (NLP) and linguistics communities, as they allow the training of models and corpus-based studies of languages. However, corpora do not exist for many languages and language varieties, such as Acadian French. In this paper, we first show that off-the-shelf NLP systems perform more poorly on Acadian French than on standard French. An Acadian French corpus could, therefore, potentially be used to improve NLP models for this dialect. Then, leveraging web-as-corpus methodologies, specifically BootCaT, domain crawling, and social media scraping, we create three corpora of Acadian French. To evaluate these corpora, drawing on the linguistic literature on Acadian French, we propose 22 statistical corpus-based measures of the extent to which a corpus is Acadian French. We use these measures to compare these newly built corpora to known Acadian French text and find that all three corpora include some traces of Acadian French.

In this paper, we consider novel cross-lingual settings for multiword expression (MWE) identification (Ramisch et al., 2020) and idiomaticity prediction (Tayyar Madabushi et al., 2022) in which systems are tested on languages that are unseen during training. Our findings indicate that pre-trained multilingual language models are able to learn knowledge about MWEs and idiomaticity that is not languagespecific. Moreover, we find that training data from other languages can be leveraged to give improvements over monolingual models.

Conventional approaches to learning word embeddings (Mikolov et al., 2013; Pennington et al., 2014) are limited to relatively few languages with sufficiently large training corpora. To address this limitation, we propose an alternative approach to deriving word embeddings for Wolastoqey and Mi’kmaq that leverages definitions from a bilingual dictionary. More specifically, following Bear and Cook (2022), we experiment with encoding English definitions of Wolastoqey and Mi’kmaq words into vector representations using English sequence representation models. For this, we consider using and finetuning sentence-RoBERTa models (Reimers and Gurevych, 2019). We evaluate our word embeddings using a similar methodology to that of Bear and Cook using evaluations based on word classification, clustering and reverse dictionary search. We additionally construct word embeddings for higher-resource languages English, German and Spanishusing our methods and evaluate our embeddings on existing word-similarity datasets. Our findings indicate that our word embedding methods can be used to produce meaningful vector representations for low-resource languages such as Wolastoqey and Mi’kmaq and for higher-resource languages.

Finite-state approaches to morphological analysis have been shown to improve the performance of natural language processing systems for polysynthetic languages, in-which words are generally composed of many morphemes, for tasks such as language modelling (Schwartz et al., 2020). However, finite-state morphological analyzers are expensive to construct and require expert knowledge of a language’s structure. Currently, there is no broad-coverage finite-state model of morphology for Wolastoqey, also known as Passamaquoddy-Maliseet, an endangered low-resource Algonquian language. As this is the case, in this paper, we investigate using two unsupervised models, MorphAGram and Morfessor, to obtain morphological segmentations for Wolastoqey. We train MorphAGram and Morfessor models on a small corpus of Wolastoqey words and evaluate using two an notated datasets. Our results indicate that MorphAGram outperforms Morfessor for morphological segmentation of Wolastoqey.

Word embeddings (Mikolov et al., 2013; Pennington et al., 2014) have been used to bolster the performance of natural language processing systems in a wide variety of tasks, including information retrieval (Roy et al., 2018) and machine translation (Qi et al., 2018). However, approaches to learning word embeddings typically require large corpora of running text to learn high quality representations. For many languages, such resources are unavailable. This is the case for Wolastoqey, also known as Passamaquoddy-Maliseet, an endangered low-resource Indigenous language. As there exist no large corpora of running text for Wolastoqey, in this paper, we leverage a bilingual dictionary to learn Wolastoqey word embeddings by encoding their corresponding English definitions into vector representations using pretrained English word and sequence representation models. Specifically, we consider representations based on pretrained word2vec (Mikolov et al., 2013), RoBERTa (Liu et al., 2019) and sentence-BERT (Reimers and Gurevych, 2019) models. We evaluate these embeddings in word prediction tasks focused on part-of-speech, animacy, and transitivity; semantic clustering; and reverse dictionary search. In all evaluations we demonstrate that approaches using these embeddings outperform task-specific baselines, without requiring any language-specific training or fine-tuning.

In this paper, we present three supervised systems for English lexical complexity prediction of single and multiword expressions for SemEval-2021 Task 1. We explore the use of statistical baseline features, masked language models, and character-level encoders to predict the complexity of a target token in context. Our best system combines information from these three sources. The results indicate that information from masked language models and character-level encoders can be combined to improve lexical complexity prediction.

Definition modelling is the task of automatically generating a dictionary-style definition given a target word. In this paper, we consider cross-lingual definition generation. Specifically, we generate English definitions for Wolastoqey (Malecite-Passamaquoddy) words. Wolastoqey is an endangered, low-resource polysynthetic language. We hypothesize that sub-word representations based on byte pair encoding (Sennrich et al., 2016) can be leveraged to represent morphologically-complex Wolastoqey words and overcome the challenge of not having large corpora available for training. Our experimental results demonstrate that this approach outperforms baseline methods in terms of BLEU score. 

Authors of text tend to predominantly use a single sense for a lemma that can differ among different authors. This might not be captured with an author-agnostic word sense disambiguation (WSD) model that was trained on multiple authors. Our work finds that WordNet’s first senses, the predominant senses of our dataset’s genre, and the predominant senses of an author can all be different and therefore, author-agnostic models could perform well over the entire dataset, but poorly on individual authors. In this work, we explore methods for personalizing WSD models by tailoring existing state-of-the-art models toward an individual by exploiting the author’s sense distributions. We propose a novel WSD dataset and show that personalizing a WSD system with knowledge of an author’s sense distributions or predominant senses can greatly increase its performance.

Cross-lingual word embeddings provide a way for information to be transferred between languages. In this paper we evaluate an extension of a joint training approach to learning cross-lingual embeddings that incorporates sub-word information during training. This method could be particularly well-suited to lower-resource and morphologically-rich languages because it can be trained on modest size monolingual corpora, and is able to represent out-of-vocabulary words (OOVs). We consider bilingual lexicon induction, including an evaluation focused on OOVs. We find that this method achieves improvements over previous approaches, particularly for OOVs.

Potentially idiomatic expressions (PIEs) are ambiguous between non-compositional idiomatic interpretations and transparent literal interpretations. For example, “hit the road” can have an idiomatic meaning corresponding to ‘start a journey’ or have a literal interpretation. In this paper we propose a supervised model based on contextualized embeddings for predicting whether usages of PIEs are idiomatic or literal. We consider monolingual experiments for English and Russian, and show that the proposed model outperforms previous approaches, including in the case that the model is tested on instances of PIE types that were not observed during training. We then consider cross-lingual experiments in which the model is trained on PIE instances in one language, English or Russian, and tested on the other language. We find that the model outperforms baselines in this setting. These findings suggest that contextualized embeddings are able to learn representations that encode knowledge of idiomaticity that is not restricted to specific expressions, nor to a specific language.

In this paper, we propose a novel method for learning cross-lingual word embeddings, that incorporates sub-word information during training, and is able to learn high-quality embeddings from modest amounts of monolingual data and a bilingual lexicon. This method could be particularly well-suited to learning cross-lingual embeddings for lower-resource, morphologically-rich languages, enabling knowledge to be transferred from rich- to lower-resource languages. We evaluate our proposed approach simulating lower-resource languages for bilingual lexicon induction, monolingual word similarity, and document classification. Our results indicate that incorporating sub-word information indeed leads to improvements, and in the case of document classification, performance better than, or on par with, strong benchmark approaches.

In this work, we consider the problem of personalizing language models, that is, building language models that are tailored to the writing style of an individual. Because training language models requires a large amount of text, and individuals do not necessarily possess a large corpus of their writing that could be used for training, approaches to personalizing language models must be able to rely on only a small amount of text from any one user. In this work, we compare three approaches to personalizing a language model that was trained on a large background corpus using a relatively small amount of text from an individual user. We evaluate these approaches using perplexity, as well as two measures based on next word prediction for smartphone soft keyboards. Our results show that when only a small amount of user-specific text is available, an approach based on priming gives the most improvement, while when larger amounts of user-specific text are available, an approach based on language model interpolation performs best. We carry out further experiments to show that these approaches to personalization outperform language model adaptation based on demographic factors.

Cross-lingual word embeddings create a shared space for embeddings in two languages, and enable knowledge to be transferred between languages for tasks such as bilingual lexicon induction. One problem, however, is out-of-vocabulary (OOV) words, for which no embeddings are available. This is particularly problematic for low-resource and morphologically-rich languages, which often have relatively high OOV rates. Approaches to learning sub-word embeddings have been proposed to address the problem of OOV words, but most prior work has not considered sub-word embeddings in cross-lingual models. In this paper, we consider whether sub-word embeddings can be leveraged to form cross-lingual embeddings for OOV words. Specifically, we consider a novel bilingual lexicon induction task focused on OOV words, for language pairs covering several language families. Our results indicate that cross-lingual representations for OOV words can indeed be formed from sub-word embeddings, including in the case of a truly low-resource morphologically-rich language.

Mi’kmaq is an Indigenous language spoken primarily in Eastern Canada. It is polysynthetic and low-resource. In this paper we consider a range of n-gram and RNN language models for Mi’kmaq. We find that an RNN language model, initialized with pre-trained fastText embeddings, performs best, highlighting the importance of sub-word information for Mi’kmaq language modelling. We further consider approaches to language modelling that incorporate cross-lingual word embeddings, but do not see improvements with these models. Finally we consider language models that operate over segmentations produced by SentencePiece — which include sub-word units as tokens — as opposed to word-level models. We see improvements for this approach over word-level language models, again indicating that sub-word modelling is important for Mi’kmaq language modelling.

In this paper we apply a range of approaches to language modeling – including word-level n-gram and neural language models, and character-level neural language models – to the problem of detecting hate speech and offensive language. Our findings indicate that language models are able to capture knowledge of whether text is hateful or offensive. However, our findings also indicate that more conventional approaches to text classification often perform similarly or better.

Verb-noun combinations (VNCs) - e.g., blow the whistle, hit the roof, and see stars - are a common type of English idiom that are ambiguous with literal usages. In this paper we propose and evaluate models for classifying VNC usages as idiomatic or literal, based on a variety of approaches to forming distributed representations. Our results show that a model based on averaging word embeddings performs on par with, or better than, a previously-proposed approach based on skip-thoughts. Idiomatic usages of VNCs are known to exhibit lexico-syntactic fixedness. We further incorporate this information into our models, demonstrating that this rich linguistic knowledge is complementary to the information carried by distributed representations.

In this paper, we propose the first model for multiword expression (MWE) compositionality prediction based on character-level neural network language models. Experimental results on two kinds of MWEs (noun compounds and verb-particle constructions) and two languages (English and German) suggest that character-level neural network language models capture knowledge of multiword expression compositionality, in particular for English noun compounds and the particle component of English verb-particle constructions. In contrast to many other approaches to MWE compositionality prediction, this character-level approach does not require token-level identification of MWEs in a training corpus, and can potentially predict the compositionality of out-of-vocabulary MWEs.

In this paper we present three unsupervised models for capturing discriminative attributes based on information from word embeddings, WordNet, and sentence-level word co-occurrence frequency. We show that, of these approaches, the simple approach based on word co-occurrence performs best. We further consider supervised and unsupervised approaches to combining information from these models, but these approaches do not improve on the word co-occurrence model.

Multiword expressions (MWEs) are lexical items that can be decomposed into multiple component words, but have properties that are unpredictable with respect to their component words. In this paper we propose the first deep learning models for token-level identification of MWEs. Specifically, we consider a layered feedforward network, a recurrent neural network, and convolutional neural networks. In experimental results we show that convolutional neural networks are able to outperform the previous state-of-the-art for MWE identification, with a convolutional neural network with three hidden layers giving the best performance.

Usage similarity (USim) is an approach to determining word meaning in context that does not rely on a sense inventory. Instead, pairs of usages of a target lemma are rated on a scale. In this paper we propose unsupervised approaches to USim based on embeddings for words, contexts, and sentences, and achieve state-of-the-art results over two USim datasets. We further consider supervised approaches to USim, and find that although they outperform unsupervised approaches, they are unable to generalize to lemmas that are unseen in the training data.

Web corpora are often constructed automatically, and their contents are therefore often not well understood. One technique for assessing the composition of such a web corpus is to empirically measure its similarity to a reference corpus whose composition is known. In this paper we evaluate a number of measures of corpus similarity, including a method based on topic modelling which has not been previously evaluated for this task. To evaluate these methods we use known-similarity corpora that have been previously used for this purpose, as well as a number of newly-constructed known-similarity corpora targeting differences in genre, topic, time, and region. Our findings indicate that, overall, the topic modelling approach did not improve on a chi-square method that had previously been found to work well for measuring corpus similarity.

In this paper we consider the problem of out-of-vocabulary term classification in web forum text from the automotive domain. We develop a set of nine domain- and application-specific categories for out-of-vocabulary terms. We then propose a supervised approach to classify out-of-vocabulary terms according to these categories, drawing on features based on word embeddings, and linguistic knowledge of common properties of out-of-vocabulary terms. We show that the features based on word embeddings are particularly informative for this task. The categories that we predict could serve as a preliminary, automatically-generated source of lexical knowledge about out-of-vocabulary terms. Furthermore, we show that this approach can be adapted to give a semi-automated method for identifying out-of-vocabulary terms of a particular category, automotive named entities, that is of particular interest to us.

Much previous research on multiword expressions (MWEs) has focused on the token- and type-level tasks of MWE identification and extraction, respectively. Such studies typically target known prevalent MWE types in a given language. This paper describes the first attempt to learn the MWE inventory of a “surprise” language for which we have no explicit prior knowledge of MWE patterns, certainly no annotated MWE data, and not even a parallel corpus. Our proposed model is trained on a treebank with MWE relations of a source language, and can be applied to the monolingual corpus of the surprise language to identify its MWE construction types.

The meanings of words are not fixed but in fact undergo change, with new word senses arising and established senses taking on new aspects of meaning or falling out of usage. Two types of semantic change are amelioration and pejoration; in these processes a word sense changes to become more positive or negative, respectively. In this first computational study of amelioration and pejoration we adapt a web-based method for determining semantic orientation to the task of identifying ameliorations and pejorations in corpora from differing time periods. We evaluate our proposed method on a small dataset of known historical ameliorations and pejorations, and find it to perform better than a random baseline. Since this test dataset is small, we conduct a further evaluation on artificial examples of amelioration and pejoration, and again find evidence that our proposed method is able to identify changes in semantic orientation. Finally, we conduct a preliminary evaluation in which we apply our methods to the task of finding words which have recently undergone amelioration or pejoration.