Um framework baseado em conhecimento de senso comum para sistemas de perguntas e respostas sobre grafo de conhecimento

Abstract

The differences between unstructured natural language and structured data make it challenging to build a computational system that supports the use of Natural Language to query Knowledge Graphs. An alternative to this problem is to use QA systems to perform queries to make this connection between data and questions in natural language. Therefore, these systems are accessible to users who do not have technical knowledge in accessing information in Knowledge Graphs, thus eliminating the need to learn complex query languages and schemas. However, using these systems presents some problems, mainly limitations in the variations observed in the natural language and the strong dependence on human interference. This work proposes a complete framework for querying Knowledge Graphs in natural language based on templates built with the help of common sense knowledge. This framework is composed of a QA system based on templates and a dataset of KGQA, along with a set of algorithms and techniques in each module of the system that can use in other approaches. Finally, ExQuestions, a question and answer dataset with multiple questions paraphrased using common sense knowledge. The main differences of our framework are: (a) how a common-sense knowledge base is used to improve the variability of the generated dataset and (b) the use of a Open IE system to improve the quality of the generated templates. Furthermore, we propose an indexing technique to retrieve the answer to the question submitted to the QA system and an experimental methodology to verify the quality of this type of system. The results obtained in this work provide contributions to the state-of-the-art of QA systems on KBQA and, finally, we make notes for future research.