Construction and Experimental Exploration of Domain Knowledge Graph for Exploratory Search
[Purpose/Significance]Information retrieval systems in specific domains often suffer from low retrieval efficiency due to the users'lack of domain knowledge.Existing domain-specific knowledge organization tools,such as domain knowledge graphs,can effectively alleviate this problem,but how to better integrate them with existing information retrieval systems is a problem that has not been solved yet.Therefore,this paper proposes a full-process solution for implementing an exploratory search system based on domain knowledge graphs,include-ing joint extraction of entities and their relations,domain knowledge graph construction and implementation of ex-ploratory search system.[Method/Process]This paper focused on the academic literature in library and information science(LIS)and explored the formal definition of entities and their relations in LIS.It also proposed a joint ex-traction method of entities and relations on Paddle UIE,and built a prototype retrieval system on an interactive LIS knowledge graph.[Result/Conclusion]By comparing the effects of different joint extraction methods of entities and relations,including Paddle UIE,CasRel,SpERT and CORE,it finds that the pre-trained large language model Paddle UIE based on prompt learning has better extraction performance,especially in the case of small sample size.In addition,this paper further designs two search experiment tasks to analyze the effect of exploratory search sys-tem.Compared with general search engines,this system can significantly improve user satisfaction and effectively solve the problem of low retrieval performance due to insufficient user domain knowledge(such as cross-disci-plinary scenarios),indicating that the full-process solution proposed in this paper can provide process guidance and suggestions for academic literature service providers to develop user support tools.
domain knowledge graphprompt learninglarge language modeljoint extractionexplorato-ry search
万方数据
维普
