A hereditary attentive template-based approach for complex knowledge base question answering systems

Abstract

Question Answering (QA) systems retrieve the most relevant answer to a natural language question. Knowledge Base Question Answering (KBQA) systems explore entities and relations from Knowledge Bases (KB) to generate answers. KBQA systems need to deal with questions that can be divided into two groups: simple and complex questions. Simple questions are those that contain direct answers that need to be detected to answer a question. Complex questions need more information than the explicit features that can be extracted from simple questions. It is necessary to use advanced query operations to collect the answer from the KB, such as exploiting indirect relations among entities, multi-relations, qualitative and quantitative constraints, and others. Currently, KBQA systems achieve better results when answering simple questions, and Complex Knowledge Base Question Answering (C-KBQA) systems turned the goal to the recent research. However, there is a lack of studies that address complex questions in the KBQA field. This work aims to fill this gap by presenting a study on C-KBQA systems. The contribution of this master thesis is twofold: a systematic mapping of the C-KBQA literature and a novel template matching approach for C-KBQA systems. First, the systematic mapping showed that C-KBQA systems need to handle with two question types: multi-hop and constraint questions. Also, it was possible to identify three main steps to construct a C-KBQA system and the use of two main approaches in this process. Secondly, our proposed C-KBQA approach performs a template matching using the combination of Semantic Parsing and Neural Networks techniques to predict the appropriate answer template to a natural language question. The so-called Hereditary Attention was created to assist the Tree-LSTM, and we demonstrate the effectiveness of our solution by comparing it to the state-of-the-art in the LC-QuAD dataset. The results show that our approach outperforms the state-of-the-art systems.