Question Answering Over Knowledge Base Method Based On Mutil-Angle Cross-Attention And Feature Enhancement

The knowledge base contains a large amount of knowledge.How to mine the answers to natural language questions from the knowledge base with a specific structure has become a research hotspot in recent years.Question answering over knowledge base mainly studies how to transform questions into structured queries to retrieve knowledge in the knowledge base.With the development of deep learning,the query method gradually evolved to combine the question and the candidate answer information in the same semantic space.After the information is compared for semantic similarity,the candidate answer corresponding to the candidate answer information with the highest semantic similarity is obtained as the answer to the question.In question answering over knowledge base research,entities in the knowledge base are usually selected as candidate answers,and multi-view information such as the path and context related to the candidate answers in the knowledge base is selected directly represent the candidate answer information,ignoring the correlation and integrity between multi-view information;when obtaining the mutual influence between the word-level question and the candidate answer information of the multi-view level,the mutual influence between the two at the overall level is ignored.Therefore,this paper proposes a multi-angle cross-attention model,which aims to use the attention mechanism from multiple angles such as question and candidate answer information to enhance the representation of the two.First,the self-attention mechanism is used to enhance the contextual information of the question and candidate answers information,and then the cross-attention mechanism is used between the multi-view information to obtain the cross-influence between the information to strengthen the representation of the candidate answer information,after the question and the candidate answer information are finally expressed as a whole,the bidirectional cross-attention mechanism is used to obtain the mutual influence of the two from the overall perspective to strengthen the expression.In addition,there are type-inspired information related to the candidate answer type and hidden temporal information related to the temporal reasoning of the candidate answer in the question,which has a certain impact on the filtering and sorting of the candidate answer.Because the information of the problem is too redundant,it is difficult for the model to obtain the key information from it.Therefore,this paper uses the feature enhancement method to strengthen the type-inspired features and hidden temporal features in the problem,through the combination of the dependency tree and the type discovery table to obtain the type discovery words in the problem,and obtain the hidden temporal words in the problem according to the temporal vocabulary table,and then use the attention mechanism to enhance the question's type-inspired features and hidden temporal features,thereby improving the filtering and ranking of candidate answers.This paper conducted experiments on the FreeBase knowledge base and WebQuestions data set.The F1 value reached 55.91%,which is better than the recent better methods.It effectively improves the accuracy of the knowledge base question and answer.And through ablation experiments,it proves that the method proposed in this paper is effective.