A flexible compositional approach to word sense disambiguation

• Autor(a) principal: Alex de Paula Barros
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/SLSC-BBKGTM
• topic: Natural Language Processing; Word Sense Disambiguation; Recuperação da informação; Computação; Processamento de linguagem natural (Computação)

Word sense disambiguation is identifying which sense of a word is used in a sentence when the word has multiple meanings. Supervised machine learning methods in which a classifier is trained for each distinct word on a corpus of manually sense-annotated examples have been the most successful algorithms to date. One possible drawback is their lack of flexibility due to requiring annotated examples for every word in the vocabulary. In contrast, knowledge-based methods do not require a classifier for each distinct word and are often built over lexico-semantic resources like ontologies, thesaurus or machine-readable dictionaries. In this work, we propose a flexible compositional algorithm based on context-gloss comparisons, that compares local context of a word represented by its neighbor words with glosses of the possible senses a word can assume using a semantic distance measure. The algorithm has three components, each based on a different information source: (i) sense frequency, obtained by counting the number of times a word occurs with each meaning in an annotated corpus, (ii) extended gloss, obtained by expanding a word dictionary definition using related words in an ontology (e.g., car and automobile), and (iii) sense usage examples, obtained from inventories that provide sentences with usage examples for some senses. Our compositional approach is flexible in the sense that it is not dependent on annotated examples and works well even when some or all of the three aforementioned knowledge sources are not available. We evaluated the performance of our algorithm for all possible combinations of the three components, simulating different scenarios of knowledge sources availability. The algorithm achieves an F1 score of 67.5 when all components are available, presenting a favorable result when compared with a state-of-the-art knowledge-based system that achieves an F1 score of 66.4