Effectiveness Analysis of Word Sense Disambiguation Using Example of Word Senses from WordNet

Effectiveness Analysis of Word Sense Disambiguation Using Examples of Word Senses from WordNet Hiroshi Sekiya Minoru Sasaki Graduate School of Sci. and Eng., Ibaraki Univ. / 4-12-1 Nakanarisawacho, Hitachi City, Ibaraki Prefecture, 316-8511, Japan Graduate School of Sci. and Eng., Ibaraki Univ. / 4-12-1 Nakanarisawacho, Hitachi City, Ibaraki Prefecture, 316-8511, Japan 22nm727x@vc.ibaraki.ac.jp minoru.sasaki.01@vc.ibaraki.ac .jp Abstract In Word Sense Disambiguation (WSD), several studies have proposed systems that incorporate dictionary glosses. The use of glosses expressed in short sentences can improve the accuracy of the WSD system. However, since many glosses are short sentences, additional lexical information could improve accuracy. In this study, we propose a method to incorporate examples of word senses described in WordNet 3.0 as new lexical information into BEM, a WSD system, and analyze the effectiveness of the examples of word senses. Specifically, examples of word senses are input into BEM, and the [CLS] vector or target word vector is taken from the output word embedding representation sequence and used as the sense embedding representation. In the experiment, out of six evaluation sets, including the development set, the F1 score decreased in five evaluation sets and improved in one evaluation set. Since the F1 score improved in one evaluation set, we expect that the use of examples of word senses would be effective. 1 Introduction Word sense disambiguation (WSD) is one of the major tasks in natural language processing (NLP). WSD is the process of identifying the most appropriate sense for a polysemous word in context. This technique is crucial in many applications in other areas of NLP, such as machine translation (Nguyen et al., 2018), information extraction (Chai and Biermann, 1999), text summarization (Rahman and Borah, 2020), and so on. Previous research has shown that incorporating lexical information, such as glosses, into a WSD system improves accuracy (Luo et al., 2018a, b; Blevins and Zettlemoyer, 2020). Glosses have been found to be effective for both most frequent sense (MFS) and less frequent sense (LFS). However, many of the glosses are written in short sentences, and it is not clear whether the glosses effectively capture the information in the sense. We expect that the use of information on many senses, rather than just glosses, will capture the characteristics of senses. To solve this problem, we propose a WSD method to use examples of word senses from WordNet 3.0 (Miller, 1995) as additional lexical information in BEM (Blevins and Zettlemoyer, 2020). In the proposed system, the target word vector or [CLS] vector of examples of word senses is used as the sense embedding representations. We expect that the use of examples of word senses as well as glosses can effectively capture the characteristics of word senses. We compare the performance of this proposed method with that of BEM to test the effectiveness of examples of word senses in WSD. 2 Related Work There are two types of word sense disambiguation (WSD) tasks: the Lexical Sample Task, in which the target words for WSD are predefined, and the All- words …

Content truncated to 500 words