Research On Single-fact Knowledge Base Question Answering Based On Multi-aspect Attention Mechanism

The knowledge base question answering has received extensive attention in recent years and is an important natural language processing task,which aims to understand and analyse the semantic meaning of a natural language problem,and then querying the answers from a structured large-scale knowledge base.In the Internet,most of the problems are based on the single-fact.After a single reasoning in the knowledge base,the answer can be obtained.Such questions are the basis of a complex question answering system.The existing research methods for single-fact knowledge base question answering are mainly based on semantic analysis and deep learning.The method based on semantic parsing relies on high-quality dictionaries and templates,which is not flexible enough.The method based on deep learning can automatically learn complex semantic information through neural networks,but the existing methods used relatively simple attention mechanism and had insufficient use of external semantic knowledge.Aimed at these problems,this paper breaks down the task into three successive steps,entity extraction,relation ranking,and answer generation.Firstly,in the entity extraction step,this paper proposes a stacked GRU network SGCLP which combines language model and part-ofspeech features.It introduces external semantic knowledge into the stacked bi-direction GRU network to improve the accuracy of entity extraction.Then,in the relation ranking step,this paper proposes a ranking model MARM that combines the multi-aspect attention mechanism to carry out multi-level interaction between the problem and the candidate relation in order to eliminate the semantic gap.After that,in the answer generation step,this paper proposes an entity ranking model RBERM that fuses relation embeddings.It uses the relation information in the knowledge base to distinguish the entities whose names are the same.Finally,the above three steps are combined to form a complete knowledge base question answering method.The experimental results show that the proposed method can effectively analyze and model the problem and query the answer through the knowledge base.On the SimpleQuestions dataset,the method of this paper achieved 78.9% accuracy based on the FB2 M knowledge base,which is close to the current best result of 80.2%.In addition,based on the FB5 M knowledge base,this paper achieved the highest accuracy of 78.3%.