Word Sense Disambiguation with Knowledge-Enhanced and Local Self-Attention-based Extractive Sense Comprehension

Proceedings of the 29th International Conference on Computational Linguistics, pages 4061–4070 October 12–17, 2022. 4061 Word Sense Disambiguation with Knowledge-Enhanced and Local Self-Attention-based Extractive Sense Comprehension Guobiao Zhang1 , Wenpeng Lu1,∗, Xueping Peng2 , Shoujin Wang3 , Baoshuo Kan1 , Rui Yu1 1School of Computer, Qilu University of Technology (Shandong Academy of Sciences), Jinan, China 2Australian Artificial Intelligence Institute, University of Technology Sydney, Sydney, Australia 3Data Science Institute, University of Technology Sydney, Sydney, Australia guobiao.zhang@foxmail.com, lwp@qlu.edu.cn {xueping.peng,shoujin.wang}@uts.edu.au 10431200583@stu.qlu.edu.cn, rui.yu1996@foxmail.com Abstract Word sense disambiguation (WSD), identify- ing the most suitable meaning of ambiguous words in the given contexts according to a pre- defined sense inventory, is one of the most classical and challenging tasks in natural lan- guage processing. Reformulating WSD as a text span extraction task is an effective ap- proach, which accepts a sentence context of an ambiguous word together with all definitions of its candidate senses simultaneously, and re- quires to extract the text span corresponding with the right sense. However, the approach merely depends on a short definition to learn sense representation, which neglects abundant semantic knowledge from related senses and leads to data-inefficient learning and subopti- mal WSD performance. To address the limita- tions, we propose a novel WSD method with Knowledge-Enhanced and Local self-attention- based Extractive Sense Comprehension (KE- LESC). Specifically, a knowledge-enhanced method is proposed to enrich semantic repre- sentation by incorporating additional examples and definitions of the related senses in Word- Net. Then, in order to avoid the huge comput- ing complexity induced by the additional in- formation, a local self-attention mechanism is utilized to constrain attention to be local, which allows longer input texts without large-scale computing burdens. Extensive experimental re- sults demonstrate that KELESC achieves better performance than baseline models on public benchmark datasets.1 1 Introduction Word sense disambiguation (WSD) is to identify a proper sense with an ambiguous word in a given context according to a predefined sense inventory, which is one of the most typical and challenging tasks in natural language processing (NLP) and ∗Corresponding author 1The source code of this paper can be obtained from https://github.com/Stubborn-z/KELESC play a critical role for human language understand- ing (Conia and Navigli, 2021). For instance, the noun word plant conveys different senses in indus- trial plant and plant seeds. WSD has been able to determine accurate meanings of ambiguous words, which is beneficial to a variety of downstream NLP applications, such as machine translation, infor- mation extraction and retrieval (Song et al., 2021; Pasini and Navigli, 2020). In recent years, with the rapid development of deep learning, the performance of WSD with neural networks-based methods has great improvement. The early neural networks-based models have cast WSD as a multi-label classification task, which disambiguated all polysemous words with a uni- fied classier (Kågebäck and Salomonsson, 2016; Raganato et al., 2017a). However, these models have focused on modeling contexts containing am- biguous words from sense-labeled training data, which ignored the rich semantic knowledge in lexi- cal resources, such as WordNet and BabelNet (Nav- igli …

Content truncated to 500 words