Abstract Meaning Representation Parsing for the Brazilian Portuguese Language

• Ano de defesa: 2020
• Autor(a) principal: Anchiêta, Rafael Torres
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Abstract meaning representation; Analisador semântico; Anotação semântica; Representação abstrata de significado; Semantic annotation; Semantic parsnig
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-29072020-120805/

Computational semantics is the area in charge of studying possible meaning representations, that is, computationally viable semantic formalisms to represent human expressions. Such formalisms play an important role in making sense of natural language, capturing the meaning of linguistic statements. Moreover, these formalisms are the main component to develop semantic parsers, which are responsible to map sentences of a natural language into a computationally treatable meaning representation. In order to represent and understand semantic features of a natural language and, with that, develop computational tools that produce results close to those of humans, several semantic formalisms were proposed, as Universal Networking Language (UNL), Universal Conceptual Cognitive Annotation (UCCA), Abstract Meaning Representation (AMR), among others. In special, AMR is a rooted directed graph-based semantic formalism with labeled nodes and edges. The nodes are concepts (that may be the words of a sentence) and the edges are semantic relations among them, where the nodes do not have an explicit alignment with the tokens of the sentences. Furthermore, AMR encompasses some linguistic features, as named entities, coreference, semantic roles, word sense disambiguation, and others. In this work, we focused on AMR representation for Portuguese, since it has a simpler structure to produce than other semantic formalisms. In this way, we annotated the Little Prince book, which is the first annotated corpus with AMR information for Portuguese and developed the first AMR parser for Portuguese. Moreover, we adapted some AMR parsing methods from English to Portuguese. More than that, we developed a new alignment strategy to align the word tokens of the sentence and the nodes of the AMR graph that improves the results of the adapted AMR parsers and a new metric to evaluate AMR graphs, which is more robust, faster, and fairer than the traditional AMR metric. Finally, we used these resources and methods in a paraphrase detection task, joining both explicit and implicit semantic features to classify if two sentences are paraphrase each other.