| The purpose of relation extraction is to extract useful relationship facts from natural language texts for downstream tasks.However,traditional supervised methods require costly manual annotation data.In recent years,the distant supervision method that automatically labels data using knowledge bases has received attention from researchers.However,the data labeled using distant supervision methods not only contains a large amount of noise but also follows a long-tail distribution.Therefore,how to train a model that can handle noisy labels and long-tail relations is currently the focus of research in the field of relation extraction.This essay addresses the issues faced by distant supervision methods by utilizing entity types and relationship alias information from knowledge bases to handle noisy labels.It also improves the multi-instance learning framework to enhance the model’s noise resistance.On this basis,the essay explicitly models the constraints between entity types and relationship labels and combines few-shot learning methods to generate constrained prototypes to improve the accuracy of long-tail relationship extraction.The main research contents are as follows:(1)Knowledge matching embedding for distantly supervised relation extraction.This essay proposes a knowledge matching embedding for distant supervision relation extraction method(KME),which addresses the problems of insufficient fusion of knowledge information and the existence of noise bags in the multi-instance learning framework.In contrast to existing methods that insufficient knowledge integration and do not consider the noise bags,KME first adds gating units to the word vectors to dynamically select position and entity type information in a matching embedding manner.Then,the multi-instance learning framework is improved by grouping sentence bags with the same relationship label and performing the final classification task on a group-by-group basis.To reduce noise,relationship alias information is introduced in the sentence-level attention mechanism to derive relationship-matching bag representations,and a similarity-based bag-level attention mechanism is used to derive group representations.Experimental results show that this method effectively reduces the impact of noise labels and improves the accuracy of relation extraction on multiple datasets.(2)Knowledge constraint prototype for distantly supervised relation extraction.The improved multiple-instance learning framework used for handling long-tailed relations can lead to an increased proportion of noisy samples,which can affect the accuracy of long-tailed relation extraction.To address the problem of noisy and few samples in long-tailed relations,this essay proposes a knowledge constraint prototype for distant supervision relation extraction method(KCP).Firstly,Trans E is used to construct paired type embeddings with structured information,and the matching embedding method is used to integrate word embeddings,introducing constraints on relation labels for paired entity types to handle noise.Then,paired type embeddings are aligned with sentence embeddings using gate units to generate sentence representations with constraint effects.Finally,a sentence-level attention mechanism based on relation matching is used to reduce the influence of noisy instances and derive knowledge-constrained prototypes for classification.Experimental results show that this method can not only handle noisy labels but also improve the performance of long-tailed relation extraction.Figure [20] table [16] reference [68]... |
