Research On Auto-answering For Non-factoid Questions

Question answering(QA)is a technique that automatically returns the corresponding accurate answer via analyzing the question posed by the user,and comprehending knowl-edge base or context information.According to the types of questions asked by users,they are roughly classified into two categories:factoid questions(such as When,Who,Where,etc.)answering and non-factoid questions(such as How,Why,Explanation-type,etc.)an-swering.This paper focuses on non-factoid questions answering(called non-factoid QA)task.For this task,the mainstream work focuses on machine comprehension reading non-factoid QA and answer selection non-factoid QA.However,QA task based on comprehen-sion reading is not only focused on non-factoid questions.Most of the existing data sets are mixed with various types of questions,which is not suitable for comprehension reading non-factoid QA task,such as Why-type QA or How-type QA;the latter work mainly introduces deep learning techniques to mine the semantic relation between the question and candidate answer,but the context information related to the question is ignored.(1)To solve the semantic inconsistency problem of answers generated by the existing Why-type QA model,this paper proposes a Why-type QA model based on causal knowledge base and passage self-matching mechanism.This model integrates multi-level attention mechanisms into sequence-to-sequence model,so as to generate more compact answer.(2)In this paper,we propose a answer filtering and answer selection based non-factoid QA framework.On the one hand,the semantic similarity method is used to filter out the irrelevant candidate answers with regard to the non-factoid question.On the other hand,we introduce answer context information from the web.As a result,the original answer selection based non-factoid QA task is transferred into a kind of complicated multi-choice comprehension reading task based on(non-factoid question,passage,candidate answers)triples.Therefore,this paper novelly proposes a non-factoid QA model based on answer context and capsule network(called FACN).This model integrates capsule network into answer-aware attention fusion layer to better model the relationship representations amongst triples,such that the best one from candidate answers is selected.This paper uses some real-world data sets to validate the effectiveness of our proposed model and method.The experimental results show that our proposed Why-type QA model based on passage self-matching technique is far superior to the existing models on both automatically-constructed Chinese and public English data sets.FACN model proposed in this paper significantly outperforms comparative answer selection non-factoid QA models on two non-factoid QA data sets.