A resolução de anáforas pronominais da língua portuguesa com base no algoritmo de Mitkov
Abstract
One of the problems of natural language processing systems is to assure
referential cohesion in a text. This property allows connecting the text constituents and
making it readable. We address the anaphoric phenomenon as one of the main factors of
referential cohesion. Anaphors depict a reference relationship between two or more text
components, and the interpretation of the anaphor is dependent upon the interpretation
of its antecedent. This work is limited to pronominal anaphors, thus, to automatic
pronoun resolution. Several algorithms have been proposed to this end. They usually
involve (1) identifying the anaphoric component; (2) determining the set of its possible
antecedents; and (3) identifying and selecting the most likely antecedent of the anaphor.
The lack of anaphora resolution in, e.g., information extraction and automatic
translation or summarization may yield non-cohesive texts. Herein we present an
adaptation of the Mitkov´s algorithm for pronoun resolution. 3rd person pronouns for
Brazilian Portuguese are especially addressed, whose antecedents are noun phrases.
This approach has been intrinsically evaluated on annotated corpora. It has also been
compared with Lappin and Leass algorithm for pronoun resolution, adapted to
Portuguese. Annotations embed morphological, syntactic and co-referential information.
The evaluation measure adopted was the success rate. This is defined as the ratio
between the number of anaphors correctly resolved by the system and the total number
of anaphors in the text. The results of both evaluations are discussed here.