Modelagem e identificação automática de construções de estrutura argumental: uma proposta para o constructicon da Framenet Brasil

Abstract

This work aims, firstly, to provide the linguistic-computational modeling, in the Constructicon of FrameNet Brazil, of 24 Brazilian Portuguese argument structure constructions, namely: Subject Predicate, Active Direct Transitive, Indirect Transitive, Oblique Transitive, Ditransitive, Intransitive, Ergative , Resultative, Object Interrupted, Split Object Artifact, Split Object Body Part, Split Object Alienable Possession, Split Object Entity Attribute, Attributive Nominal Predicative, Stative Nominal Predicative, Locative Predicative, Objective Nominal Verb, State Change, Without Subject Intransitive, Existential Presentation and Elapsed Time. Second, it presents and discusses the results of an automatic construction identification task whose objective is to identify instances of the above constructions. In this sense, the proposal fits into the theoretical-methodological discussions of FrameNet Brazil (FNBr) (SALOMÃO, 2009a), which develops two major computational resources: a Lexicon and a Constructicon. Lexicon aims to: create a computational representation of frames, defined by its participants and instruments, and connected to each other via relationships between frames; define Lexical Units, pairings between a lemma and a meaning defined in terms of a frame; note sentences that exemplify the syntactic and semantic valence patterns in which ULs occur. Constructicon, on the other hand, consists of a computational resource in which structural information about constructions of a language is modeled and it is built through a semantic model based on frames, linked to its linguistic manifestations (SALOMÃO, 2009a). Both resources are interconnected, as both Lexicon and Constructicon were designed to deal with all the linguistic features relevant to the identification of a frame and a construction. In view of the interconnection between the two fronts of the FN-Br, we will present a proposal that intends to expand the scope of previous research (ALMEIDA, 2016), to assist in Natural Language Comprehension tasks. Through the interaction between Linguistics and Computing, this work encompasses the three phases presented by Dias da Silva (2006) as necessary for the development of a system in Computational Linguistics: the linguistic phase, the linguisticcomputational phase and the computational phase. Therefore, it presents the specification of constructions in the linguistic phase, the representation of these constructions in the Constructicon, in the linguistic-computational phase, and proposes a computational system so that these constructions can be recognized in the computational phase. This system aims to test the hypothesis that the modeling of constructions, composed of information related to their form and meaning, helps in the task of recognizing constructs licensed by them. For that, instances of the constructions were annotated in corpus and submitted to different models of Artificial Intelligence to be classified according to the existing constructions in Constructicon. The results indicated that, for very frequent constructions in the training corpus, the addition of information from the modeling negatively affects the performance of the model. On the other hand, for less frequent constructional patterns, such modeling positively affects the performance of the model