SIM: um modelo semântico inferencialista para expressão e raciocínio em sistemas de linguagem natural

Abstract

Computer processing of natural language is one of the major challenges in the area of Artificial Intelligence. There is an urgent need for computational systems that can help us process the overwhelming quantity of non-structured information that is present in natural language. Driven by this need, we studied various philosophies of language in order to understand what a linguistic expression consists of, or at least what reponses (borrowed from Philosophy) can make it possible for computer systems to address the needs and challenges in an effective way. For representationalist theories, concepts are things (or the state of affairs) represented by words and linguistic expressions. Therefore, there is the idea of a single and regular world—an inductive world. Pragmatic philosophies, such as that of Wittgenstein, define a new concept of "concept": the content of a concept consists of its various uses in language games. Concepts can be defined and understood based on the practice of applying concepts in situations of language use. Inferentialist philosophers such as Sellars, Dummett and Brandom incorporate the project of a semantic theory by presenting a reduction of the pragmatic approach to the uses of concepts in rational games, where the inferential component is the most important. In the second part of this research, we study linguistic resources and systems in the fields of Computational Linguistics and Natural Language Processing (NLP) in order to define a conceptual framework for the way that ordinary systems express and compute meanings. The model of the Computational Semantics—a traditional model in NLP—proposes that (i) one representation of the world is sufficient to define the semantic value of terms and sentences in natural language; and (ii) there is equivalence between the syntax and the semantics of natural languages. As a result of our research, we propose a new computational model for expression and semantic reasoning in natural language systems—the Semantic Inferentialism Model (SIM). SIM is based on Robert Brandom's Semantic Inferentialism Theory. We advocate the idea that in order for a computational system to answer questions, extract and retrieve information, refute arguments, justify answers, or give explanations about a text, among many other applications, it must: (i) express the inferential content of concepts and sentences (pre-conditions and post-conditions of use); and (ii) manipulate this inferential content within the flow of reasoning. The semantic bases of SIM were constructed and have formed the first large-scale linguistic resource for the Portuguese language with inferentialist content—InferenceNet.BR. The semantic reasoning algorithm (SIA) and the InferenceNet.BR resource were applied in a real system for extracting information on crime—the WikiCrimes Information Extractor (WikiCrimesTE). The results of the extraction were evaluated and superseded the state of the art of the foremost systems that execute tasks of understanding natural language, even for the English language. The greatest legacy of this research was to have planted a seed for research in Inferentialist Computational Semantics, which we believe will allow a more solid and fruitful evolutionary path for the field of Computational Linguistics.