A Generative Causal Question Answering Method Based On Automatic Knowledge Base Construction

Question answering is an important task with valuable applications in artificial intelligence.Various challenging tasks have been introduced for the development of question answering.Gaokao,as China's national college entrance exam,requires not only strong abilities of natural language processing and reasoning,but also rich knowledge of the relevant domains.This thesis presents the first attempt to resolve the problem of causal essay questions in the geography exams of Gaokao which need long answer generation.For the needed domain knowledge,this thesis suggests a description of event knowledge and relations in a form of abstract event graph,and design a automatic extraction and construction process of abstract event graph based on a corpus of question-answer pairs.For question understanding,answer generation and reasoning,this thesis proposes a graph-based multi-hop reasoning question answering model,which unifies unstructured information of questions and structured information of graphs.After the deep understanding process,a decoder in the model generates answers.BERT,graph neural networks and transformer decoder are utilized to accomplish these functions.The approach that this thesis represents significantly outperforms a variety of question answering methods,which shows its effectiveness.