Non-Parametric Few-Shot Learning for Word Sense Disambiguation

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1774–1781 June 6–11, 2021. ©2021 Association for Computational Linguistics 1774 Non-Parametric Few-Shot Learning for Word Sense Disambiguation Howard Chen, Mengzhou Xia, Danqi Chen Princeton University {howardchen, mengzhou, danqic}@cs.princeton.edu Abstract Word sense disambiguation (WSD) is a long- standing problem in natural language process- ing. One significant challenge in supervised all-words WSD is to classify among senses for a majority of words that lie in the long- tail distribution. For instance, 84% of the an- notated words have less than 10 examples in the SemCor training data. This issue is more pronounced as the imbalance occurs in both word and sense distributions. In this work, we propose MetricWSD, a non-parametric few-shot learning approach to mitigate this data imbalance issue. By learning to com- pute distances among the senses of a given word through episodic training, MetricWSD transfers knowledge (a learned metric space) from high-frequency words to infrequent ones. MetricWSD constructs the training episodes tailored to word frequencies and explicitly addresses the problem of the skewed distri- bution, as opposed to mixing all the words trained with parametric models in previous work. Without resorting to any lexical re- sources, MetricWSD obtains strong perfor- mance against parametric alternatives, achiev- ing a 75.1 F1 score on the unified WSD evalu- ation benchmark (Raganato et al., 2017b). Our analysis further validates that infrequent words and senses enjoy significant improvement.1 1 Introduction Word sense disambiguation (WSD) (Navigli, 2009) is a widely studied problem that aims to assign words in text to their correct senses. Despite ad- vances over the years, a major challenge remains to be the naturally present data imbalance issue. Mod- els suffer from extreme data imbalance, rendering learning the long-tail examples a major focus. In the English all-words WSD task (Raganato et al., 2017b), 84% of the annotated words2 have less 1Our code is publicly available at: https://github. com/princeton-nlp/metric-wsd. 2Here we use “word” for simplicity. In WSD datasets, a word is a combination of its stem and part-of-speech tag. 101 102 103 104 0 250 500 # Words Word Frequency 0.0 0.2 0.4 0.6 0.8 1.0 0 100 200 # Words % of MFS Examples Figure 1: Top: the long-tail distribution of words in the training data (SemCor). Bottom: percentage of exam- ples whose target senses are the most frequent sense for a given word (only words with ≥ 10 examples are considered). All single-sense words are excluded. than 10 occurrences in the training data and the most frequent sense (MFS) accounts for a large portion of the examples, resulting in a 65.2 test F1 score by simply predicting MFS (Figure 1). Recent approaches tackle this problem by resort- ing to extra sense information such as gloss (sense definition) and semantic relations to mitigate the issue of rare words and senses (Luo et al., 2018b,a; Kumar et al., 2019; Huang et al., 2019; Blevins and Zettlemoyer, 2020; Bevilacqua and Navigli, 2020). However, most work …

Content truncated to 500 words