State-of-the-art pretrained contextualized models (PCM) eg.BERT use tasks such as WiC and WSD to evaluate their wordin-context representations.This inherently assumes that performance in these tasks reflect how well a model represents the coupled word and context semantics.We question this assumption by presenting the first quantitative analysis on the context-word interaction being tested in major contextual lexical semantic tasks.To achieve this, we run probing baselines on masked input, and propose measures to calculate and visualize the degree of context or word biases in existing datasets.The analysis was performed on both models and humans.Our findings demonstrate that models are usually not being tested for word-in-context semantics in the same way as humans are in these tasks, which helps us better understand the model-human gap.Specifically, to PCMs, most existing datasets fall into the extreme ends (the retrieval-based tasks exhibit strong target word bias while WiC-style tasks and WSD show strong context bias); In comparison, humans are less biased and achieve much better performance when both word and context are available than with masked input.We recommend our framework for understanding and controlling these biases for model interpretation and future task design.

Capturing word meaning in context and distinguishing between correspondences and variations across languages is key to building successful multilingual and cross-lingual text representation models.However, existing multilingual evaluation datasets that evaluate lexical semantics "in-context" have various limitations.In particular, 1) their language coverage is restricted to high-resource languages and skewed in favor of only a few language families and areas, 2) a design that makes the task solvable via superficial cues, which results in artificially inflated (and sometimes super-human) performances of pretrained encoders, and 3) no support for crosslingual evaluation.In order to address these gaps, we present AM 2 ICO (Adversarial and Multilingual Meaning in Context), a widecoverage cross-lingual and multilingual evaluation set; it aims to faithfully assess the ability of state-of-the-art (SotA) representation models to understand the identity of word meaning in cross-lingual contexts for 14 language pairs.We conduct a series of experiments in a wide range of setups and demonstrate the challenging nature of AM 2 ICO.The results reveal that current SotA pretrained encoders substantially lag behind human performance, and the largest gaps are observed for low-resource languages and languages dissimilar to English.

Abstract Research into representation learning models of lexical semantics usually utilizes some form of intrinsic evaluation to ensure that the learned representations reflect human semantic judgments. Lexical semantic similarity estimation is a widely used evaluation method, but efforts have typically focused on pairwise judgments of words in isolation, or are limited to specific contexts and lexical stimuli. There are limitations with these approaches that either do not provide any context for judgments, and thereby ignore ambiguity, or provide very specific sentential contexts that cannot then be used to generate a larger lexical resource. Furthermore, similarity between more than two items is not considered. We provide a full description and analysis of our recently proposed methodology for large-scale data set construction that produces a semantic classification of a large sample of verbs in the first phase, as well as multi-way similarity judgments made within the resultant semantic classes in the second phase. The methodology uses a spatial multi-arrangement approach proposed in the field of cognitive neuroscience for capturing multi-way similarity judgments of visual stimuli. We have adapted this method to handle polysemous linguistic stimuli and much larger samples than previous work. We specifically target verbs, but the method can equally be applied to other parts of speech. We perform cluster analysis on the data from the first phase and demonstrate how this might be useful in the construction of a comprehensive verb resource. We also analyze the semantic information captured by the second phase and discuss the potential of the spatially induced similarity judgments to better reflect human notions of word similarity. We demonstrate how the resultant data set can be used for fine-grained analyses and evaluation of representation learning models on the intrinsic tasks of semantic clustering and semantic similarity. In particular, we find that stronger static word embedding methods still outperform lexical representations emerging from more recent pre-training methods, both on word-level similarity and clustering. Moreover, thanks to the data set’s vast coverage, we are able to compare the benefits of specializing vector representations for a particular type of external knowledge by evaluating FrameNet- and VerbNet-retrofitted models on specific semantic domains such as “Heat” or “Motion.”

We present the first evaluation of the applicability of a spatial arrangement method (SpAM) to a typologically diverse language sample, and its potential to produce semantic evaluation resources to support multilingual NLP, with a focus on verb semantics.We demonstrate SpAM's utility in allowing for quick bottom-up creation of large-scale evaluation datasets that balance cross-lingual alignment with language specificity.Starting from a shared sample of 825 English verbs, translated into Chinese, Japanese, Finnish, Polish, and Italian, we apply a two-phase annotation process which produces (i) semantic verb classes and (ii) fine-grained similarity scores for nearly 130 thousand verb pairs.We use the two types of verb data to (a) examine cross-lingual similarities and variation, and (b) evaluate the capacity of static and contextualised representation models to accurately reflect verb semantics, contrasting the performance of large language-specific pretraining models with their multilingual equivalent on semantic clustering and lexical similarity, across different domains of verb meaning.We release the data from both phases as a large-scale multilingual resource, comprising 85 verb classes and nearly 130k pairwise similarity scores, offering a wealth of possibilities for further evaluation and research on multilingual verb semantics.

One of the most powerful features of contextualized models is their dynamic embeddings for words in context, leading to state-of-the-art representations for context-aware lexical semantics. In this paper, we present a post-processing technique that enhances these representations by learning a transformation through static anchors. Our method requires only another pre-trained model and no labeled data is needed. We show consistent improvement in a range of benchmark tasks that test contextual variations of meaning both across different usages of a word and across different words as they are used in context. We demonstrate that while the original contextual representations can be improved by another embedding space from both contextualized and static models, the static embeddings, which have lower computational requirements, provide the most gains.

We present a novel methodology for fast bottom-up creation of large-scale semantic similarity resources to support development and evaluation of NLP systems. Our work targets verb similarity, but the methodology is equally applicable to other parts of speech. Our approach circumvents the bottleneck of slow and expensive manual development of lexical resources by leveraging semantic intuitions of native speakers and adapting a spatial multi-arrangement approach from cognitive neuroscience, used before only with visual stimuli, to lexical stimuli. Our approach critically obtains judgments of word similarity in the context of a set of related words, rather than of word pairs in isolation. We also handle lexical ambiguity as a natural consequence of a two-phase process where verbs are placed in broad semantic classes prior to the fine-grained spatial similarity judgments. Our proposed design produces a large-scale verb resource comprising 17 relatedness-based classes and a verb similarity dataset containing similarity scores for 29,721 unique verb pairs and 825 target verbs, which we release with this paper.

In this paper, we present a thorough investigation on methods that align pre-trained contextualized embeddings into shared crosslingual context-aware embedding space, providing strong reference benchmarks for future context-aware crosslingual models.We propose a novel and challenging task, Bilingual Token-level Sense Retrieval (BTSR).It specifically evaluates the accurate alignment of words with the same meaning in crosslingual non-parallel contexts, currently not evaluated by existing tasks such as Bilingual Contextual Word Similarity and Sentence Retrieval.We show how the proposed BTSR task highlights the merits of different alignment methods.In particular, we find that using context average type-level alignment is effective in transferring monolingual contextualized embeddings cross-lingually especially in non-parallel contexts, and at the same time improves the monolingual space.Furthermore, aligning independently trained models yields better performance than aligning multilingual embeddings with shared vocabulary.

Word sense disambiguation and the related field of automated word sense induction traditionally assume that the occurrences of a lemma can be partitioned into senses. But this seems to be a much easier task for some lemmas than others. Our work builds on recent work that proposes describing word meaning in a graded fashion rather than through a strict partition into senses; in this article we argue that not all lemmas may need the more complex graded analysis, depending on their partitionability. Although there is plenty of evidence from previous studies and from the linguistics literature that there is a spectrum of partitionability of word meanings, this is the first attempt to measure the phenomenon and to couple the machine learning literature on clusterability with word usage data used in computational linguistics. We propose to operationalize partitionability as clusterability, a measure of how easy the occurrences of a lemma are to cluster. We test two ways of measuring clusterability: (1) existing measures from the machine learning literature that aim to measure the goodness of optimal k-means clusterings, and (2) the idea that if a lemma is more clusterable, two clusterings based on two different “views” of the same data points will be more congruent. The two views that we use are two different sets of manually constructed lexical substitutes for the target lemma, on the one hand monolingual paraphrases, and on the other hand translations. We apply automatic clustering to the manual annotations. We use manual annotations because we want the representations of the instances that we cluster to be as informative and “clean” as possible. We show that when we control for polysemy, our measures of clusterability tend to correlate with partitionability, in particular some of the type-(1) clusterability measures, and that these measures outperform a baseline that relies on the amount of overlap in a soft clustering.

Unsupervised word sense disambiguation (WSD) methods are an attractive approach to all-words WSD due to their non-reliance on expensive annotated data.Unsupervised estimates of sense frequency have been shown to be very useful for WSD due to the skewed nature of word sense distributions.This paper presents a fully unsupervised topic modelling-based approach to sense frequency estimation, which is highly portable to different corpora and sense inventories, in being applicable to any part of speech, and not requiring a hierarchical sense inventory, parsing or parallel text.We demonstrate the effectiveness of the method over the tasks of predominant sense learning and sense distribution acquisition, and also the novel tasks of detecting senses which aren't attested in the corpus, and identifying novel senses in the corpus which aren't captured in the sense inventory.

Word sense disambiguation (WSD) is an old and important task in computational linguistics that still remains challenging, to machines as well as to human annotators. Recently there have been several proposals for representing word meaning in context that diverge from the traditional use of a single best sense for each occurrence. They represent word meaning in context through multiple paraphrases, as points in vector space, or as distributions over latent senses. New methods of evaluating and comparing these different representations are needed. In this paper we propose two novel annotation schemes that characterize word meaning in context in a graded fashion. In WSsim annotation, the applicability of each dictionary sense is rated on an ordinal scale. Usim annotation directly rates the similarity of pairs of usages of the same lemma, again on a scale. We find that the novel annotation schemes show good inter-annotator agreement, as well as a strong correlation with traditional single-sense annotation and with annotation of multiple lexical paraphrases. Annotators make use of the whole ordinal scale, and give very fine-grained judgments that “mix and match” senses for each individual usage. We also find that the Usim ratings obey the triangle inequality, justifying models that treat usage similarity as metric. There has recently been much work on grouping senses into coarse-grained groups. We demonstrate that graded WSsim and Usim ratings can be used to analyze existing coarse-grained sense groupings to identify sense groups that may not match intuitions of untrained native speakers. In the course of the comparison, we also show that the WSsim ratings are not subsumed by any static sense grouping.

There has been a great deal of recent research into word sense disambiguation, particularly since the inception of the Senseval evaluation exercises. Because a word often has more than one meaning, resolving word sense ambiguity could benefit applications that need some level of semantic interpretation of language input. A major problem is that the accuracy of word sense disambiguation systems is strongly dependent on the quantity of manually sense-tagged data available, and even the best systems, when tagging every word token in a document, perform little better than a simple heuristic that guesses the first, or predominant, sense of a word in all contexts. The success of this heuristic is due to the skewed nature of word sense distributions. Data for the heuristic can come from either dictionaries or a sample of sense-tagged data. However, there is a limited supply of the latter, and the sense distributions and predominant sense of a word can depend on the domain or source of a document. (The first sense of “star” for example would be different in the popular press and scientific journals). In this article, we expand on a previously proposed method for determining the predominant sense of a word automatically from raw text. We look at a number of different data sources and parameterizations of the method, using evaluation results and error analyses to identify where the method performs well and also where it does not. In particular, we find that the method does not work as well for verbs and adverbs as nouns and adjectives, but produces more accurate predominant sense information than the widely used SemCor corpus for nouns with low coverage in that corpus. We further show that the method is able to adapt successfully to domains when using domain specific corpora as input and where the input can either be hand-labeled for domain or automatically classified.

In word sense disambiguation (WSD), the heuristic of choosing the most common sense is extremely powerful because the distribution of the senses of a word is often skewed. The problem with using the predominant, or first sense heuristic, aside from the fact that it does not take surrounding context into account, is that it assumes some quantity of hand-tagged data. Whilst there are a few hand-tagged corpora available for some languages, one would expect the frequency distribution of the senses of words, particularly topical words, to depend on the genre and domain of the text under consideration. We present work on the use of a thesaurus acquired from raw textual corpora and the WordNet similarity package to find predominant noun senses automatically. The acquired predominant senses give a precision of 64% on the nouns of the SENSEVAL-2 English all-words task. This is a very promising result given that our method does not require any hand-tagged text, such as SemCor. Furthermore, we demonstrate that our method discovers appropriate predominant senses for words from two domain-specific corpora.

Selectional preferences have been used by word sense disambiguation (WSD) systems as one source of disambiguating information. We evaluate WSD using selectional preferences acquired for English adjective—noun, subject, and direct object grammatical relationships with respect to a standard test corpus. The selectional preferences are specific to verb or adjective classes, rather than individual word forms, so they can be used to disambiguate the co-occurring adjectives and verbs, rather than just the nominal argument heads. We also investigate use of the one-senseper-discourse heuristic to propagate a sense tag for a word to other occurrences of the same word within the current document in order to increase coverage. Although the preferences perform well in comparison with other unsupervised WSD systems on the same corpus, the results show that for many applications, further knowledge sources would be required to achieve an adequate level of accuracy and coverage. In addition to quantifying performance, we analyze the results to investigate the situations in which the selectional preferences achieve the best precision and in which the one-sense-per-discourse heuristic increases performance.

Research into the automatic acquisition of subcategorization frames (SCFs) from corpora is starting to produce large-scale computational lexicons which include valuable frequency information. However, the accuracy of the resulting lexicons shows room for improvement. One significant source of error lies in the statistical filtering used by some researchers to remove noise from automatically acquired subcategorization frames. In this paper, we compare three different approaches to filtering out spurious hypotheses. Two hypothesis tests perform poorly, compared to filtering frames on the basis of relative frequency. We discuss reasons for this and consider directions for future research.