Joint Extraction Of Nested Named Entity And Relations Based On Multi-task Learning

Nested named entity recognition and relation extraction are important tasks of information extraction.They are key parts of high-level applications and tasks such as information retrieval,question answering system,and intelligent recommendation in natural language processing systems.However,the existing research pays less attention to the task of nested entity recognition,and cannot make full use of the rich semantic relationships between entities.At the same time,traditional methods often regard nested named entity recognition and entity relationship extraction as independent tasks.The method has the problem of cascading errors,and it cannot make full use of the associated information between the two tasks.Moreover,there is a lack of labeling strategies and data sets for the joint extraction of nested entities and relation.In response to the above problems,this article has done the following works:1.We propose a relation labeling strategy based on hierarchical structure.In this paper,the extraction of nested entity relation is defined as a sequence labeling problem.Aiming at the cross-layer problem in the labeling of nested entity relation,the relationship labeling is divided into same-level entity relation labeling and cross-layer relation labeling,and labeling methods are designed respectively.We use the labeling strategy designed by the article to improve the entity and relationship labeling of Genia,KBP and NYT datasets.2.We propose a joint extraction model of nested named entities and relations based on two-level structure.The model uses the method of parameter sharing to realize the joint learning of the two tasks.The correlation feature of the two tasks improves the effectiveness of the joint model.At the same time,we designed a dual dynamic hierarchical network structure.We insert multiple rounds of hierarchical relationship extraction on the basis of hierarchical nested named entity recognition.The method effectively solves the problem of crosslayer entity relationship extraction under the hierarchical recognition of nested named entities.3.Multi-group control experiments and ablation experiments were done on the three datasets.Experiments show that the joint extraction model proposed in this paper has improved effects on both tasks.On the task of nested named entity recognition,the F1 index has an average improvement of 4.3%.In Relation Extraction tasks,the F1 index has an average improvement of 2%.