Know Your Graph. State-of-the-Art Knowledge-Based WSD

Proceedings of Recent Advances in Natural Language Processing, pages 949–958, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_110 949 Know Your Graph. State-of-the-Art Knowledge-Based WSD Alexander Popov and Kiril Simov and Petya Osenova IICT, Bulgarian Academy of Sciences Sofia, Bulgaria {alex.popov|kivs|petya}@bultreebank.org Abstract This paper introduces several improve- ments over the current state of the art in knowledge-based word sense disam- biguation. Those innovations are the re- sult of modifying and enriching a knowl- edge base created originally on the basis of WordNet. They reflect several sepa- rate but connected strategies: manipulat- ing the shape and the content of the knowl- edge base, assigning weights over the rela- tions in the knowledge base, and the ad- dition of new relations to it. The main contribution of the paper is to demonstrate that the previously proposed knowledge bases organize linguistic and world knowl- edge suboptimally for the task of word sense disambiguation. In doing so, the pa- per also establishes a new state of the art for knowledge-based approaches. Its best models are competitive in the broader con- text of supervised systems as well. 1 Introduction Word sense disambiguation (WSD) is a long- standing task in natural language processing (NLP), which has been approached through several broad families of computational tech- niques. While supervised learning models typi- cally achieve the highest accuracy scores, prob- lems related to the lack of gold corpora, data sparseness and suboptimal granularity of the com- putational lexicons have plagued the field and prevented significant breakthroughs. This paper presents results achieved with knowledge-based word sense disambiguation (KBWSD) algorithms as an alternative pathway. It builds on previ- ous work in the subfield and demonstrates that KBWSD can achieve accuracy scores near and in some cases even at the state-of-the-art, as a rule dominated by supervised approaches. More- over, the experimental results indicate that in- depth analysis of knowledge representation and knowledge enrichment hold significant promises – both as an alternative to supervised WSD, able to sidestep data-related issues, and as a source of potentially powerful new training signals. The contribution of the article is fourfold. Firstly, it presents a novel strategy for construct- ing knowledge bases used in KBWSD, showing that a structure centered on word senses rather than on synsets can be more effective for this partic- ular task. Secondly, another novel approach to structuring the graph is explored in using word and synset embedding models in order to assign weights to relation arcs in the knowledge graph. Thirdly, a number of avenues for the enrichment of the semantic network used for KBWSD are pur- sued, thus linking WordNet to external resources. Finally, a new state of the art for KBWSD is estab- lished, which is competitive even when compared with supervised systems. The paper is structured as follows: the next sec- tion presents related work; section 3 deals with some preliminary experiments aimed at replicat- ing previous work described in the literature, but in a slightly modified setting; section 4 outlines several strategies for improving the structure of the …

Content truncated to 500 words