Research And Application Of Question Answering Methods Based On Semantic Parsing And Multi-Hop Reasoning

With the advent of the era of information intelligence,the amount of information in the Internet has exploded.The user used to input a question in the search engine,and the search engine retrieve the certain web page from a large amount of web pages on the Internet.Although it can also give the desired answer,it requires the user to find the answer from the retrieval results.This greatly reduces the efficiency and accuracy of information acquisition.Based on the above problems,the question answer system is in urgent need.Question answering systems are gradually favored by researchers,and are widely used in finance,medical care,Internet,marketing and other fields.There are three main types of question answering systems currently under study.The question answering system based on common questions cannot cover all the questions of users and has great limitations.Knowledge graph based methods can give more accurate answers,but it is time and human labor intensive to build a knowledge graph.The construction cost of question answering system based on unstructured knowledge is low,but the system accuracy rate is not sufficient.To sum up,a question answering system based on a single method cannot obtain multiple forms of existing knowledge in the specific field,and it is difficult to balance the effect and cost.Therefore,it is necessary to study question answering systems based on multiple methods for different knowledge structures.This paper investigates semantic parsing based knowledge base question answering and multi-hop reasoning based machine comprehension reading question answering systems.Finally,an intelligent question answering system based on the two question answer methods is constructed,and its final output is computed through the question and answer fusion decision module.The main research work is as follows:(1)Research on knowledge base question answering algorithm based on semantic parsing.The knowledge base question answering method based on semantic parsing is a method to obtain the question answer from the knowledge base through entity relationship linking and constructing formal query.Aiming to lower the noise that may be introduced into candidate queries in the process of constructing formal queries,this paper studies the method for generating formal queries,and proposes a method for generating and sorting abstract query directed graphs.The method generates abstract query directed graph structure according to the question,and then generates candidate query graphs based on the abstract query directed graph.The method can reduce the noise introduced in the process of generating the candidate query graph,and improves the accuracy of the results in the subsequent candidate query graph sorting task.The experimental results show that the formal query generation method proposed in this paper can significantly reduce the number of generated candidate query graphs and improve the precision value,the recall value and F1 value of the candidate query graph ranking task.(2)Research on question answering algorithm for machine reading comprehension based on multi-hop reasoning.Machine reading comprehension question answering is a method of obtaining answers to questions from documents by retrieving clues,extracting spans,and reasoning paths.Aiming at the problem that clues obtained by machine reading comprehension in the retrieval stage are not enough to support the subsequent multi-hop reasoning task,which leads to the reduction of the accuracy of question answering,this paper studies the introduction method of external knowledge base,and extracts the clues related to the question from the knowledge base.The proposed filtering module removes paragraphs which are not related to the question,and obtains more paragraphs related to the question.Based on the retrieved paragraphs,the final answer to the question is obtained through a span extraction module,a multi-hop reasoning module and an answer selection module.The experimental results show that the machine reading comprehension method which introduces the external knowledge base can retrieve more clues related to the question,and improves the exact matching value and F1 value of support facts extraction and answer extraction tasks.(3)Use structured and unstructured data in the automotive field to design and develop an intelligent question answering system integrating two question answering methods.The system includes a knowledge graph question answering module,a machine reading comprehension question answer module,and a final question answering fusion module.The knowledge graph question answering module and the machine reading comprehension module process the questions raised by users in parallel to output the answers,and the question answering fusion module selects the final answer through a decision strategy.Based on the algorithm model studied in this paper,the questions related to the field of automobile pre-sales and after-sales are answered.Through the functional and non-functional test of the system,it is proved that this system in the automobile pre-sale and after-sale field integrating two question and answering algorithms can correctly answer the questions raised.