Estudo e validação de teorias do domínio lingüístico com vistas à melhoria do tratamento de cadeias de co-referência em sumarização automática

Abstract

The work presented in the dissertation focuses on the study and validation of linguistic theories so as to improve reference cohesion in Automatic Summarization systems, which with the advent of the Internet have received increasing attention due to the urge to manage the huge amounts of on-line textual information that become available each day. In this dissertation we evaluate Seno (2005) s Veins Theory-based proposal and prototype, and present a reimplementation with distinct features based on the analysis of a corpus annotated with rhetoric (RST) and referential information. In addition, we report on the first validation effort for Portuguese for Veins Theory s Conjecture 1 (C1), which constrains anaphora resolution given the rhetoric structure of texts and whose applicability to Automatic Summarization interests us. As a methodological novelty, we put forth the Non-Trivial Precision, a more realistic estimator of C1 s predictive power.