基于多语义相似性的关系检测方法

A relation detection method based on multi semantic similarity

关系检测是知识库问答的关键步骤,直接影响问答质量。现有方法中基于编码比较的方法提取文本整体语义进行匹配会丢失序列的局部信息,而基于交互的方法在文本低层表征层面进行比较会忽略全局语义。针对现有方法无法兼顾全局语义和局部语义信息的问题,提出了一种基于多语义相似性的关系检测模型,通过BERT模型分别对问题和关系进行语义表示,然后引入注意力机制、双向长短期记忆网络和多层感知机进行局部关联性分析;利用BERT计算出的句向量中含有序列的全局语义信息,设计了问题和关系句向量的全局相似度度量。在基准数据集SimpleQuestions和WebQSP上进行了实验验证,所提方法分别取得了93.92%和87.81%的准确率,优于其他现有的方法。

Relation detection is a critical step of knowledge base question answering,which directly affects the quality of question answering.Among the existing methods,the encoding-comparison method extracts text global semantic information for matching,which often ignores the local semantic feature of text sequence.The interaction approach performs the comparison on low-level representations based on the sequence local information,which fails to consider the global semantic information of the input sequences.To solve the issues,this paper proposes a relation detection model based on Bert model and multi-semantic similarity considering global and local semantic information.First,our model introduces Bert as a text encoding layer to represent questions and relations as sequences of vectors.And then,a bi-directional long short-term memory(Bi-LSTM)layer with the attention mechanism is used to analyze the local semantic relevance and calculate the local similarity.Finally,our model uses a distance calculation formula to measure the global semantic relevance between questions and relations.The experimental results on two benchmark datasets,SimpleQuestions and WebQSP,show that the proposed model achieves the accuracy of 93.92%and 87.81%respectively,performs better than state-of-the-art approaches.