Linguistics parameters for zero anaphora resolution

Pereira, Simone Cristina

http://hdl.handle.net/10400.1/1787

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
Dissertation_2010_05_26_revisada.pdf		409.02 KB	Adobe PDF	Download

Send Feedback

Authors

Pereira, Simone Cristina

Advisor(s)

Baptista, Jorge

Evans, Richard J.

Abstract(s)

This dissertation describes and proposes a set of linguistically motivated rules for zero anaphora resolution in the context of a natural language processing chain developed for Portuguese. Some languages, like Portuguese, allow noun phrase (NP) deletion (or zeroing) in several syntactic contexts in order to avoid the redundancy that would result from repetition of previously mentioned words. The co-reference relation between the zeroed element and its antecedent (or previous mention) in the discourse is here called zero anaphora (Mitkov, 2002). In Computational Linguistics, zero anaphora resolution may be viewed as a subtask of anaphora resolution and has an essential role in various Natural Language Processing applications such as information extraction, automatic abstracting, dialog systems, machine translation and question answering. The main goal of this dissertation is to describe the grammatical rules imposing subject NP deletion and referential constraints in the Brazilian Portuguese, in order to allow a correct identification of the antecedent of the deleted subject NP. Some of these rules were then formalized into the Xerox Incremental Parser or XIP (Ait-Mokhtar et al., 2002: 121-144) in order to constitute a module of the Portuguese grammar (Mamede et al. 2010) developed at Spoken Language Laboratory (L2F). Using this rule-based approach we expected to improve the performance of the Portuguese grammar namely by producing better dependency structures with (reconstructed) zeroed NPs for the syntactic-semantic interface. Because of the complexity of the task, the scope of this dissertation had to be limited: (a) subject NP deletion; b) within sentence boundaries and (c) with an explicit antecedent; besides, (d) rules were formalized based solely on the results of the shallow parser (or chunks), that is, with minimal syntactic (and no semantic) knowledge. A corpus of different text genres was manually annotated for zero anaphors and other zero-shaped, usually indefinite, subjects. The rule-based approached is evaluated and results are presented and discussed.