Research On Chinese Knowledge Base Question Answering Based On Neural Networks

Question answering system is a popular research topic in the fields of information re-trieval,artificial intelligence and natural language processing,and has a very broad develop-ment prospect.Knowledge base question answering(KBQA)is an important task of the re-search,that is,given a natural language question,and the question answering system searches related entities or text from the existing knowledge base as the final answer to the question.In the process of understanding and answering questions,multiple natural language process-ing tasks such as question classification,entity mention recognition,entity linking,candidate relation recognition,and semantic analysis are involved.The question classification is a core part of question understanding,and the quality of the question answering system is directly affected by the classification accuracy.Candidate relation recognition is a kind of semantic matching task,calculating the semantic similarity between question and candidate relations,that is a key step in answer search.Based on the neural network models,this thesis studies the related tasks of the Chinese KBQA.The main contents are as follows:(1)Chinese Question Classification Model Based on Language Model and Attention MechanismAiming at problems such as the short length of Chinese question sentences,the lack of semantic feature information and the uncertain position of interrogative words in sentences,a new classification model is proposed in this thesis.Compared with the traditional word em-bedding model,this method uses the latest pre-trained language model to further enhance the distributed representation of each word in the question,then acquires the semantic features of the context through a bidirectional long short term memory network,and finally uses the at-tention model to strengthen attention to interrogative information in sentences.Experimental results on three datasets prove that our method can get an average performance improvement of 3.63%over the benchmark model,which reflects good classification performance.(2)Candidate Relationship Recognition Method Based on Semantic Similarity Com-putingTraditional methods mostly treat text as a set of words,construct the feature vector by counting the number of times each word appears in the text,and then use metrics such as cosine distance to calculate the text similarity.However,these methods only consider the characteristics of the word level in the sentence,not the semantic level,and ignore a lot of important information,including syntax and word order.We propose a new neural network framework.This method designs multiple attention mechanisms based on Siamese network to achieve semantic similarity matching between sentence pairs.The experimental results show that the model can take full advantage of the semantic information of the text,and the F1 value is 84.59 on the dataset provided by the CCKS2018 semantic matching evaluation task,ranking fourth.(3)Chinese Knowledge Base Question Answering Based on Multi-label StrategyAt present,many existing methods can only deal with simple questions that can be an-swered with only one triplet knowledge,but cannot solve complex questions that involve multiple triplet knowledge,including multiple entities and multiple relations.This thesis proposes a new question answering system based on multi-label strategy,which includes two main modules:question processing and answer search.In the problem processing mod-ule,different model frameworks are proposed for the three tasks of the subject entity mention recognition,entity linking and candidate relation recognition.With the help of multiple tags,the existing Chinese questions are divided into simple questions,chained questions,and multi-entity questions.In the answer search module,different solutions are designed for the above three questions.Experimental results indicate that our proposed method can achieve better performance.The average F1 value on the validation set of CCKS2019-CKBQA pub-lic evaluation data is 66.76,ranking second.