One Classifier for All Ambiguous Words: Overcoming Data Sparsity by Utilizing Sense Correlations Across Words

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5978–5985 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5978 One Classifier for All Ambiguous Words: Overcoming Data Sparsity by Utilizing Sense Correlations Across Words Prafulla Kumar Choubey, Ruihong Huang Department of Computer Science and Engineering Texas A&M University, College Station, TX 77840 prafulla.choubey@tamu.edu, huangrh@cse.tamu.edu Abstract Most supervised word sense disambiguation (WSD) systems build word-specific classifiers by leveraging labeled data. However, when using word-specific classifiers, the sparseness of annotations leads to inferior sense disambiguation performance on less frequently seen words. To combat data sparsity, we propose to learn a single model that derives sense representations and meanwhile enforces congruence between a word instance and its right sense by using both sense-annotated data and lexical resources. The model is shared across words that allows utilizing sense correlations across words, and therefore helps to transfer common disambiguation rules from annotation-rich words to annotation-lean words. Empirical evaluation on benchmark datasets shows that the proposed shared model outperforms the equivalent classifier-based models by 1.7%, 2.5% and 3.8% in F1-score when using GloVe, ELMo and BERT word embeddings respectively. Keywords: Word Sense Disambiguation, Systematic Polysemy, Lexical Semantics 1. Introduction Word sense disambiguation (WSD) aims to automatically identify the correct meaning of a word in a particular con- text and is essential for many downstream natural lan- guage processing tasks, including information extraction, text classification, information retrieval, and machine trans- lation. Most existing data-driven approaches for Word sense dis- ambiguation (WSD) build word-specific classifiers to pre- dict the right sense of a word instance, which lack the capa- bility to generalize across words and therefore require suf- ficient sense-annotated data for every word (de Lacalle and Agirre, 2015) in order to disambiguate them well. Con- sequently, the model’s performance decreases significantly when there is a lack of training data for a word or some of its senses. As shown in Table 1, the state-of-the-art word-specific classifier model (Luo et al., 2018b) is able to achieve over 90% F1 score on senses with more than 200 training instances. But the performance of the model drops quickly on annotations-lean senses, especially on senses with less than 10 training instances. Essentially, the design of word-specific classifiers overlooks the correlations among senses of different words. Studies on systematic polysemy (Apresjan, 1974; Rumshisky and Batiukova, 2008; Boleda et al., 2012) have shown that similar senses exist widely between words denoting objects of the same category, e.g., as shown in Table 2, the words “bank” and “school” can both refer to a building and an institution; similarly, the verbs “digest” and “swallow” share two related senses depending on whether the argument refers to concrete objects or abstract con- cepts. Modeling systematic polysemy (Pustejovsky, 1995; Utt and Pad´o, 2011) and accurately identifying words with similar senses is challenging though, which is not our focus here. Instead, we aim to directly address the limitation of word-specific classifiers for WSD that completely isolate a word from others …

Content truncated to 500 words