Research And Implementation Of Key Technologies On Question-answering System Based On Media Knowledge Graph

In recent years,with the explosive growth of information in the era of big data and the growing demand for knowledge in vertical domains such as media,the construction of knowledge graphs in vertical domains and its question answering system have become a research hotspot today.This thesis focuses on the shortcomings of traditional construction of vertical domain knowledge graph,such as large manual dependence,low construction efficiency,and insufficient concise answers in traditional question answering methods,focusing on key technologies such as the construction of vertical domain knowledge graphs and Chinese named entity recognition tasks in question answering models.The specific research contents are as follows:(1)Research is conducted on the method of constructing a vertical knowledge graph based on semi-automation.Aiming at the problem that the current vertical domain knowledge graph is mostly constructed manually,this thesis proposes the Semi-automated Domain Knowledge Graph Construction Method(SA-KG),which decouples the traditional highly coupled manual-based method,and at the same time,the technical steps to be intelligent are automated,thereby improving the construction efficiency.Firstly,the problem of high human involvement in the construction of domain schema is investigated,and a semi-automatic domain schema construction method is designed,which is mainly based on machine automation and complemented by manual annotation.Aiming at the problems of diversification of original data structure and information redundancy of unstructured data in the process of knowledge extraction,research on knowledge extraction methods based on multi-source heterogeneous data is carried out,and the Cas Rel model is selected to extract entity relationships from unstructured data through multiple sets of comparative experiments.Aiming at the problem of conflicts in multi-source data storage,the research on knowledge fusion technology based on similarity algorithm is carried out,and Neo4 j graph database is selected for knowledge storage.(2)Research on the question answering model based on deep learning is carried out.Aiming at the problems that the Chinese named entity recognition model,the key task of the question answering model,has the phenomenon of lack of lexical information and the lack of using domain dictionaries,which leads to the problems of inaccurate model recognition and low generalization,this thesis proposes the Soft Lexicon Lattice with Deep Multi-Mutual learning Networks(MM-SLLattice).By fully combining open dictionary information and domain dictionary information,the method introduces word information and boundary information into the input representation layer losslessly,and applies mutual learning to the model training of sequence tasks.Finally,the comparison experiment with the mainstream model is carried out,and the results show that this model is better than the mainstream model.(3)Based on the above research content,combining named entity recognition,entity linking and relationship prediction,facing the application scenario of knowledge graph in the media field,building a knowledge base in the media field,and completing a KGQA in the media field,and the basic function of the system is tested and demonstrated.