Research On Few-shot Joint Entity Relationship Extraction Based On Meta-learning

With the advent of the era of big data,the generation of massive data makes it possible to train deep and complex deep learning models.However,due to its disordered and unstructured nature,massive training data is difficult to be directly used for model training,and manual annotation is still required.A lot of time and labor costs are invested in data labeling and data processing.It has become one of the recent research hotspots to enable the model to quickly and efficiently perform well in the current task in the context of few-shot.In the field of natural language processing,how to extract valuable information from texts and how to quickly understand the semantics and focus of massive texts have always been the core topics.The maturation of knowledge graph technology has also promoted the development of deep learning tasks that use entity-relationship extraction to form entity-relationship triples.Among them,how to use the end-to-end joint model to directly complete the two tasks of named entity recognition and relationship extraction has gradually become the mainstream entity relationship extraction task solution path due to the superior performance of the model.Based on the above background,this paper mainly focuses on the task of joint entity relationship extraction under the background of few-shot learning.It is required to create a few-shot entity-relationship extraction corpus.(1)Use data crawler to collect word-of-mouth comment data of Autohome,build an entity-relationship triplet system,and create a few-shot entity-relationship extraction corpus according to the requirements of the few-shot task.(2)On the basis of the self-built few-shot auto word-of-mouth review entity relationship extraction corpus,Cas Rel cascaded pointer network is selected as the entity relationship extraction model,and the meta-learning algorithm MAML and Reptile are used as the model training algorithm to verify the model performance.Experiments show that the meta-learning algorithm that optimizes the initial features can significantly improve the model performance of the few-shot joint entity relationship extraction task;the more samples provided by a single category,the better the model performance,the more meta-training categories,the worse the model performance;Cas Rel joint entity relationship extraction model performs better;the number of iterations of the base learner of the MAML algorithm has little effect on the performance of the algorithm,while the number of iterations of the base learner of the Reptile algorithm has a significant impact on the performance of the algorithm;there is no significant difference in the performance of the MAML algorithm and the Reptile algorithm.(3)After fine-tuning other corpora of the same type and the same task,the model is used to extract entity-relationship triples,and it is concluded that car owners pay more attention to the battery life and power of electric vehicles after new energy vehicles gradually enter the market.