Modelagem construcional de anáforas pronominais na FrameNet Brasil: contribuições para o mapeamento computacional da referência

Loading...
Thumbnail Image

Journal Title

Journal ISSN

Volume Title

Publisher

Universidade Federal de Juiz de Fora (UFJF)

Abstract

The work presented in this thesis is included in the FrameNet Brasil studies (FN Br). Through the theoretical basis of Frame Semantics (FILLMORE, 1982) and Berkeley Construction Grammar (KAY & FILLMORE, 1999), it aims to (i) provide the computational and linguistic treatment required to model Brazilian Portuguese anaphorical constructions (ii) discuss the application of the Constructicon model to information extraction methods. To achieve these goals, this work proposes a linguistic-computational representation of the demonstrative, possessive, reflexive, relative, undetermined, nominative, oblique and treatment anaphorical constructions. By analyzing data in corpora, we noted some diversity related to pronominal anaphora patterns in Portuguese language. From that, we modeled the constructions according to the types of anaphora and we defined the elements which compose the structure: the antecedent and the pronoun. Furthermore, we specified all the elements which are part of the structure using linguistic descriptions – constraints based on Universal Dependencies (UDs), constituent order and frames. At last, we implemented a construction recognition system for pronominal anaphoras and their antecedents using the constructions we modeled in the constructicon tool. On the one hand, the results suggest a correct mapping of antecedents by applying the model, on the other hand they suggest some exceptions related to mapping some types of antecedents, such as complex noun phrases, named entities, ellipses or cases when they are located outside the sentence boundary. Those exceptions are related to three reasons: deficiencies related to the anaphora model itself, restrictions in the FrameNet database, which can’t process named entities, and, lastly, restrictions in the UD parser, which processes lexical items in one sentence each time. The results emphasize the need to extend the coverage of the FN-Br Constructicon.

Description

Citation

Endorsement

Review

Supplemented By

Referenced By

Creative Commons license

Except where otherwise noted, this item's license is described as Acesso Aberto