A Question Answering System Base On Terminology Mining And Siamese Neural Network

With the increase of user information acquisition requirements,traditional search engines have become increasingly difficult to meet user needs.Although the search engine can return relevant knowledge to the user,the user still needs secondary extraction.For this reason,researchers have proposed an intelligent question answering system,which can return users with an accurate answer.According to the types of answers,question answering systems are divided into two types: factoid and non-factoid.Factoid question answering systems usually take an entity or concept as the answer,and non-factoid question answering systems usually use a sentence or paragraph as the answer.Non-factoid answer can explain the problem in detail,so it is often used in domain-specific.This paper conducts research on non-factoid question answering systems in domain-specific.The non-factoid question answering system mainly obtains answers from the question answering database to respond user input.At present,most question answering systems are designed based on state-of-the-art neural network models.With the help of semantic models,appropriate answers are selected and returned to users.However,there are still many problems.For example,neural network model treats the learned semantic information equally,professional terms are just as important as normal words.This cannot reflect the importance of domain knowledge to the question answering system.Therefore,the existing non-factoid question answering systems still have difficulties in dealing with problems in specific domain.Aiming at this problem,this paper proposes a domain-specific non-factoid question answering systems by combining information retrieval technique and deep neural network.Then based on the RASA open source framework,an intelligent question answering assistant in the domain-specific was implemented.This paper mainly carried out the following work:First,we extract professional terms from the domain-specific documents.The professional terms are composed of several normal words.Due to the complexity of Chinese semantics and the limitation of word segmentation tools,this has caused many difficulties in the research of professional terms extraction.Aiming at this problem,this paper proposes a domain-specific terminology mining method that extracts terms from documents in specific domains.First,using a frequent word mining algorithm to extract a frequent word set as a complement to domain knowledge;Then,according to the frequency and part-of-speech characteristics of the terminology to filter the candidate frequent word set;finally,a professional term database is constructed after verification on the web knowledge base.Second,most question answering systems have achieved satisfied performance in the open domain,but do not perform well in domain-specific.The main reasons are: one is the lack of domain knowledge;the other is that the semantic model treats domain knowledge and common knowledge equally,and it is difficult to reflect the importance of domain knowledge.Therefore,this paper designs a non-factoid question answering system in a specific domain,which mainly includes three modules:The terminology mining module extracts professional terms from the domain-specific documents;The semantic matching module trains a semantic sentence matching model offline based on a deep siamese neural network;The answer retrieval module processes the user's online question through a query step and a ranking step.Finally,this paper conducted experiments based on two real domain-specific QA databases,and the experiment results have demonstrated the effectiveness of the proposed QAS.