Research On Multi-task-based Meta Learning Method In Fine-grained Entity Typing

Since the development of the Internet,massive amounts of unstructured information data have been generated at any time,and it needs to be automatically converted into structured knowledge data storage and utilization by using Information Extraction Technology.The Fine-grained Entity Typing task is one of the key basic tasks for the research of Information Extraction,and it provides the key technical support for the construction of Knowledge Graph and Knowledge Base.In the field of Fine-grained Entity Typing,high-quality data resources are scarce,and manual labeling costs are high,and data has become the bottleneck of the model.How to use the existing resource data to obtain better model generalization ability,and how to effectively use the newly labeled resources in a small sample area,are urgent problems to be solved.This paper proposes two approaches to the above challenges,one is a multi-task learning method for integrating existing data sets,and the other is a multitask meta-learning method for a small sample field.(1)A method of Fine-grained Entity Typing based on multitasking.This paper designs a hard parameter sharing mechanism based on multi-task learning.By integrating multiple existing data sets and self-created data sets,a universal model is obtained.This model uses the hierarchical information of entity types to construct an embedded representation of entity types;and by sharing feature extraction layers,it implicitly adds training data and improves the learning ability of the network;in the task layer,it enhances task-related information from the shared layer makes the output of the shared layer more suitable for the task itself,and further improves the model fitting ability.Experimental results show that both the new data set and the multi-task learning method have a strong improvement effect on Fine-grained Entity Typing tasks,and the best model has a 50% improvement over the original benchmark model.(2)Multi-task-based Meta Learning method for Fine-grained Entity Typing.This paper designs a multi-task-based meta learning experiment,training tests on new tasks,and then comparing the model's ability to learn.Based on the above,this paper proposes two methods: one is a meta-learning algorithm based on gradient descent,which has the characteristics of fast fitting and wide universality;the other is a meta learning algorithm based on prototype network.Establish a prototype representation for each entity type in each task,so that the model learns the prior distribution between tasks,and by using the training results of the prior distribution,the model achieves better results on the new data set.Finally,experimental results show that the results of the multi-tasking-based meta-learning model can be improved by nearly 40% on the benchmark model,further proving the superiority of the above two methods,and providing a Fine-grained Entity Typing problem in a small number of samples new ideas.The two methods proposed in this paper eliminate the gap between the existing finegrained entity classification datasets and also solve the problem of fine-grained entity classification in a small sample area to a certain extent.In addition,the method proposed in this paper participated in the TAC Knowledge Base Population 2019 and achieved the first domestic achievement.The method was directly applied to the China Engineering Science and Technology Knowledge Center Construction Project led by the Chinese Academy of Engineering,which played a role in the construction of knowledge base.Important role.