Anaphora resolution without world knowledge

• Autor(a) principal: Leffa, Vilson J.
• Data de Publicação: 2018
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: DELTA: Documentação de Estudos em Lingüística Teórica e Aplicada
• Texto Completo: https://revistas.pucsp.br/index.php/delta/article/view/38224
• topic: Anaphora Resolution; Natural Language Processing; Textual Constraints; Ambiguity; Anaphora Resolution; Natural Language Processing; Textual Constraints; Ambiguity

A typical problem in the resolution of pronominal anaphora is the presence of more than one candidate for the antecedent of the pronoun. Considering two English sentences like (1) "People buy expensive cars because they offer more status" and (2) "People buy expensive cars because they want more status" we can see that the two NPs "people" and "expensive cars", from a purely syntactic perspective, are both legitimate candidates as antecedents for the pronoun "they". This problem has been traditionally solved by using world knowledge (e.g. schema theory), where, through an internal representation of the world, we "know" that cars "offer" status and people "want" status. The assumption in this paper is that the use of world knowledge does not explain how the disambiguation process works and alternative explanations should be explored. Using a knowledge poor approach (explicit information from the text rather than implicit world knowledge) the study investigates to what extent syntactic and semantic constraints can be used to resolve anaphora. For this purpose, 1,400 examples of the word "they" were randomly selected from a corpus of 10,000,000 words of expository text in English. Antecedent candidates for each case were then analyzed and classified in terms of their syntactic functions in the sentence (subject, object, etc.) and semantic features (+ human, + animate, etc.). It was found that syntactic constraints resolved 85% of the cases. When combined with semantic constraints the resolution rate rose to 98%. The implications of the findings for Natural Language Processing are discussed.