Design And Implementation Of Vehicle Information Question Answering System Based On Deep Learning Hybrid Model

The search engine is an important deployment in the field of information retrieval.People can use search engines to find what they need to improve the efficiency of solving problems in work and life.However,with continuous development of the information era,the amount of network data has grown rapidly,people may obtain duplicated or irrelevant web pages through search engines.People expect more concise and accurate answers.The intelligent question answering system just meets people's needs.This personified intelligent system can receive questions with natural language,analyze users' intentions and then return users with direct and accurate answers through inference and calculation.Due to the continuity of Chinese sentences and the flexibility of Chinese grammar,there is not a mature open source framework of Chinese question answering system currently.In addition,the accuracy of deep learning algorithms in the question answering task still needs much more improvement.Furthermore,the prevailing question answering system is implemented in the open level,and it may not be possible to return satisfied answers to users in the professional level.Therefore,the Chinese question answering system bases on deep learning for the limited level has great research value.The type of question answering system can be classified according to the data domain to the question domain,the answer data,the feedback mechanism for getting answer,and the method used in the information retrieval.This thesis focuses on a Chinese question answering system,which based on the limited field with automotive knowledge,question and answer pairs,retrieval and deep learning methods.The purpose of this thesis is to improve the accuracy of question answering system in experiments and to implement a vehicle information question answering system of a car manual.This thesis focuses on three aspects,the construction of the limited domain knowledge base,the application of deep learning in the sentence matching,and the construction of a question answering system.The construction of the limited domain knowledge base mainly involves a lexicon in the automotive field and the replacement of re-statement,processing a car manual and constructing question and answer pairs.We choose text as the information organization form of the knowledge base.In simple terms,we obtain the question corpus of the knowledge base through web crawler,and then fill in answers corresponding to the contents of the car manual to construct the knowledge base in the automobile field.In terms of deep learning,there are some limitations in the previous sentence matching model: firstly,some studies use BiLSTM or CNN algorithm alone,without considering that algorithm fusion can represent sentence pairs more comprehensively.Secondly,some studies only consider the oneway attention mechanism in the fusion algorithm,but not use the two-way attention mechanism that can make the representation of two sentences in a sentence pair interact each other.Aiming at the limitations of the existing models,this thesis proposes a BiGRU-based hybrid model.This model combines the BiGRU algorithm which is simpler than the BiLSTM algorithm with the CNN algorithm,and adds the two-way attention mechanism to the pooling layer.Using this model to conduct experiments on open data sets to prove the effectiveness of this model in sentence matching.Based on the above two research works and combining technologies such as natural language processing,semantic analysis,artificial intelligence,and information retrieval,this thesis proposes an vehicle information question answering system to fully understand the question from a car user,and retrieve a answer corresponding to the question and answer pair with the highest similarity to the question to solve the problem.