Effective Use of Japanese Dictionary Definition Sentences in Learning Hierarchical Embedding of Dictionaries

Effective Use of Japanese Dictionary Definition Sentences in Learning Hierarchical Embedding of Dictionaries Yuki Ishii Minoru Sasaki Ibaraki University Ibaraki University 22nm702y@vc.ibaraki.ac.jp minoru.sasaki.01@vc.ibaraki.ac .jp Abstract Existing knowledge graph-based learning methods for word sense disambiguation use word-to-word relations to learn models, but do not learn using the hierarchical relations of word senses in a single word. In addition, even when defining sentences in a Japanese dictionary are applied, the accuracy of judging the hierarchical relationship of word senses is poor, and the effect of knowledge graph embedding learning is not fully achieved. This study analyzes how to edit dictionary descriptions to improve the accuracy of models that judge the hierarchical relationship between senses in a Japanese dictionary. The results of the analysis showed that the accuracy of the unedited dictionary was 60.9%, while the accuracy of the edited dictionary was 83.3%, confirming the improved performance of the model. 1 Introduction Word sense disambiguation (WSD) is the task of identifying which sense is used in a sentence for a polysemous word in the sentence. In the last few years, many knowledge graph-based approaches have been studied that do not require the cost of word sense labeling. Related works include GlossBERT (Huang et al., 2019) and BEM (Blevins and Zettlemoyer, 2020), which combine word sense definition sentences with supervised learning, and EWISE (Kumar et al., 2019) and EWISER (Bevilacqua and Navigli, 2020), which use word-to- word relations in knowledge data. However, these methods do not learn models using the hierarchical relationship of word senses in words. When the EWISE system learned the hierarchical relationship of word senses using the Japanese dictionary definition sentences in their original form, it was found that the relationship judgment accuracy of the generated models was low, and the knowledge information was not sufficiently learned. In the Iwanami Japanese Dictionary, there are definition sentences that only describe usage expressions and part-of-speech expressions such as "《attached to noun 》 " and "((adjective))". In Japanese, it is expressed as "《名詞に付けて》" and "((形))". Such definitions are likely to exist in multiple dictionaries and overlap in content, thus inhibiting learning of the hierarchical relationship between senses. Therefore, in addition to the existing relations between words, we will learn the hierarchical relations of word senses and analyze how the contents of the Japanese dictionary can be edited and modified to improve the accuracy of the hierarchical relation judgment model of word senses. Learning a model for judging the hierarchical relationship of word senses is an initial task to improve the performance of the knowledge graph embedding system. To investigate how well the information in the knowledge graph can be embedded, we use the knowledge graph embedding learning of the EWISE system, which provides accuracy in judging the relationship between triples of data in the knowledge data. 2 Related Work Many approaches based on knowledge graphs have been studied. In GlossBERT, a pre-trained BERT encoder is given a context and a word definition sentence, and is trained to judge whether the word definition sentence correctly …

Content truncated to 500 words