Research On Key Technologies Of Question Answering Based On Knowledge Graph

The essence of knowledge is interconnection.Only the knowledge based on interconnection can exert the exponential effect of data integration and reflect the semantic value of big data.The emergence of knowledge maps provides a more accurate expression for semantic information retrieval.The infrastructure of the field is widely used.It is one of the current hotspots to structure online knowledge,construct a knowledge map that facilitates the extraction of semantic information,and conduct automatic question answering research based on the knowledge map.Knowledge graph based question answering algorithm,its core is how to correctly analyze the semantics of user questions and find accurate answers in a huge knowledge base.According to the characteristics of the knowledge graph based question answering task,the whole process is decomposed into three key sub-tasks,namely entity recognition,entity link and relationship detection.In the process of entity recognition,taking into account the characteristics of the question itself,in order to address the problem of insufficient semantic representation capabilities of existing named entity recognition models,this paper introduces the latest semantic model BERT,and proposes an improved algorithm for topic word recognition based on self-encoding language models.In the Simple Questions dataset,more than 95% of the F1-scores have been achieved,which has reached the SOTA level.In the entity link task,this paper proposes a greedy search algorithm based on the popularity score of the entity,which improves the model's robustness and ranking accuracy.Solve the problem of non-alignment of traditional methods.For the relationship detection task,in view of the fact that most current research methods use vector modeling ideas to embed the facts in the question and knowledge map into a common feature space,and use the cosine similarity calculation to find the correct answer to the question.However,such methods usually have tedious steps in the selection of candidate main entities and often ignore the correlation between the main entity and the relationship in the question and the relationship between the original words.Therefore,this paper proposes the MHA-SCNN model,which uses different structural features of RNN and CNN to extract more comprehensive semantic features.The model uses multiple layers of Bi LSTM to capture deeper problem expressions and connect residuals to match the semantic and word meaning information of the relationship,and uses multiple attention mechanisms to obtain more reasonable matching representations of the relationship.At the same time,CNN similarity matrix is adopted to model the natural language word level.The experimental results show that the model proposed in this paper can achieve significant results in the Free Base knowledge graph.At the same time,from the financial field,this article explores the method of constructing the domain knowledge map for data source collection,entity relationship extraction,knowledge fusion,etc.,and builds a question answering system on this basis,using the traditional semantic analysis method to achieve the effect of accurate question answering to a certain extent.