Desenvolvimento de um modelo inteligente capaz de responder questões consultando bancos de dados estruturados (KBS)

Abstract

The virtual sales market (E-commerce) has expanded a lot nowadays due to the ease and prac- ticality provided by this purchase way, and due to the fact that the technologies are becoming

more and more accessible. With the increase in consumers who use this method of purchase,

the implementation of virtual assistants by companies can add benefits to both sides of the ne- gotiation consumers-companies, since the use of virtual assistants can allow the automation of

tasks involving means of communication, thus accelerating problem resolution and increasing customer service productivity, in addition to being able to provide personalized experiences tailored to each consumer. In this work, concepts that are relevant to the topic of Knowledge Base Question Answering (KB-QA) are discussed, and ways to develop an end-to-end KB-QA

model that is capable of answering user questions provided in natural language, based on in- formation contained in knowledge bases (KBs) that can be seen as graphs. To this end, two

parallel lines of research are addressed, complete models, which are able to directly deal with

the question in natural language and provide an answer based on the KB, and composite mo- dels, models that are composed of a method for converting the question from natural language

in a query and a method of querying the KB via queries, where the query is usually made up of triples of the form (subject, predicate, object). In order to develop an end-to-end KB-QA model via Ensemble by majority vote, superior on terms of metrics like F1-score, this work addresses the generation of a composite KB-QA model, where methods for transforming questions from natural language to queries and methods to query the Knowledge Bases based on these queries are analyzed and selected, the search and analysis of complete KB-QA models is carried out,

and then some of the models are selected for the composition of the Ensemble. In the analy- sis of the composite KB-QA model, databases from different areas such as COUNTRIES and

FB15k-237 are applied for independent analysis of the method for querying the KB, in order to determine the potential of the model for different structures of the Knowledge Bases based on triples, and then the WikiMovies database is used to analyze the whole composite model. For the analysis of the complete KB-QA models, the WebQuestionsSP question bank is used, which is developed to be answered through queries to the KB Freebase. As the main result of this work, three different Ensembles are generated, called Simple Ensemble, Ensemble with Countermeasure and Ensemble with total countermeasure, which are state-of-the-art for the KB-QA task considering the WebQuestionsSP database with F1-scores of 75.40%, 78.72% and

81.43% respectively. Besides these results, as an addendum, some analyzes and results are pre- sented involving the development of a composite KB-QA model, here called Parot-MINERVA

composite model.