Research And Implementation Of Non-specific Domain Chinese Intelligent Question Answering System Based On Deep Neural Network Model

Intelligent question answering system(IQAS)is a new type of information service system,which integrates information retrieval,natural language processing,artificial intelligence,semantic analysis and other technologies.It can make intelligent and accurate response to users’ problems.IQAS can be divided into specific areas and open field QA system in accordance with the scope of questions,and it also can be divided into factual and non-factual QA according to the question type.This paper designs and implements an open domain intelligent question answering system,uses different strategies to deal with different types of problems,focuses on the process of dealing with factoid questions,template matching techniques and deep sequence mapping model for non-factual questions.Because of the open domain,the designed template can’t cover all the non-factual questions.Therefore,this paper introduces the sequence mapping model based on the deep neural network,and uses a mountain of collected question and answer pairs corpus to train the model and learn the mapping between the sequence of questions and the sequence of answers,which is equivalent to self-learning "template" and can produce a response to the input problem.It can solve the problem of insufficient coverage of the templates.The IQAS implemented in this paper mainly includes the deep QA process of factual questions,such as question type judgment,support evidence retrieval,support evidence score,candidate answer extraction and scoring steps.The techniques involved in this paper include text preprocessing process,AIML template matching,Lucene full text search,recurrent neural network(RNN)and its deformation long and short term memory model(LSTM).The main works of the author are as follows:Firstly,we build the sequence mapping model based on RNN to map the QA pairs,change the original inputs from the sparse word bag vectors to the dense distributed representation word vectors,and design experiments to explore the impact of different parameters on model training.The experimental results show that the method is 14% higher than Lucene’s results,indicate the sequence mapping model can learn some hidden rules in the training corpus and can make reasonable responses to the unknown problems in the corpus.Secondly,we analyze and compare the present supporting evidence and candidate answer score components.We found that the candidate answer score method based on the average word distance model can use the structure information of the sentence.The answer score method based on the word frequency applies to the situation that the expected answer appears multiple times in the supporting evidence,but lacks a uniform expression pattern.Therefore,this method combines the two candidate answer score methods which takes into account the structure,word frequency is used in the system.The experiments show that the combined method has a higher accuracy and MRR value.Thirdly,we design the system framework,and finally realize the Chinese intelligent QA prototype system based on the above research results.