Research On Relation Detection For Question Answering Over Knowledge

The goal of Knowledge Base Question Answering(KBQA)is to rely on a well-built knowledge base to return precise answers to questions in natural language.As a key subtask in KBQA,Relation Detection aims to extract the relation in the question.The traditional search method can only return a collection of documents,which requires the user to further process the feedback results,while the introduction of the KBQA technology can directly return the final answer.Therefore,KBQA has become a supporting technology for practical applications such as search engines and smart speakers and has important research value.The research goal of this paper is to improve the relation detection model and validate the role of relation detection in question answering.This paper builds an overall framework for the KBQA task and considers the introduction of pre-trained language models,Recurrent Neural Networks,and attention mechanisms to improve each sub-step of the knowledge question answering framework.The research content mainly includes the following three aspects:1.Among the existing relation detection 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 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,Simple Questions and Web Questions SP,show that the proposed model achieves the accuracy of 93.92% and 87.81% respectively,performing better than state-of-the-art approaches.2.Aiming at the problem of insufficient semantic representation in the existing question answering methods based on knowledge graph embedding,a question answering method based on entity scoring model and relation detection model is proposed.Firstly,the entity scoring model uses the BERT pre-training language model to convert the question of input into a vector sequence,and the pre-trained knowledge graph embedding is used to convert entities into representation vectors.In construction,the three-layer stacked Gated Recurrent Unit is used to further extract the context information of the vector sequence,and then the vector sequence is transformed into the problem representation vector by using the attention mechanism after a Layer Norm layer.Finally,the question representation vector,topic entity representationvector,and candidate entity representation vector are input into the Compl Ex function and reserve the top-200 candidate entities.Then the relation detection step is performed,and finally,the answer is calculated by the answer selection module.Experiments are performed on the benchmark datasets Web Questions SP and Meta QA,and the results show the effectiveness of this method.3.The proposed knowledge question answering method is applied to the open domain question answering dataset,and a knowledge question answering prototype system is implemented.The prototype system provides a web interface to interact with users.