Research On Relation Extraction Method Based On Paraphrase And Multi-Information Fusion

Relation Extraction aims to extract the semantic relation between a given entity from large-scale corpus,which is the important basis of natural language applications such as knowledge base construction and question answering.Paraphrase refers to the phenomenon of homogeneity in natural language,which is used to describe sentences with similar semantics but different grammatical structures.Combing with the related techniques of paraphrase,the following optimizations are made for the lack of the corpus and information on the relation representation model.(1)Distant Supervision Relation Extraction based on Paraphrase ConstraintsIn view of the lack of training corpus in relation extraction,this paper proposes a distant supervision relation extraction based on paraphrase constraints.Firstly,the corpus is extended by distant supervision method.On this basis,the definition of the relation type is introduced to restrict the relation sample as the semantic constraint of it.This paper uses paraphrase identification method to determine whether there is a paraphrase relation between the extended sample and the definition to filter the samples that do not satisfy the paraphrase relation.The experiment proves that the performance of this method is better than the existing state-of-the-art supervised learning model,and alleviates the noise impact brought by remote supervision.(2)Relation Representation Model with Multi-granularity InformationAiming at the error accumulation problem caused by manual features in traditional relation extraction methods and the lack of representation ability of the existing deep learning model,we proposes a convolutional recurrent neural network model that combines multi-granularity information.This model combines the advantages of convolutional neural network to extract multi-granularity local features and the ability of recurrent neural network to capture sequence information.At the same time,multi-granularity features are merged through various integration strategies such as attention mechanism and element-wise max-pooling.The experimental results show that this method achieves better performance than the mainstream relation extraction system without adding any additional information.(3)Definition-based and Attention-based Multi-Instance ModelThe above-mentioned distant supervision relation extraction method based on paraphrase constraint faces the problem of "injury" in sample filtering.To solve this problem,this paper proposes a multi-instance model based on definition constraint and attention mechanism.This method uses the multi-instance learning strategy to model the different instances that contains the same entity pair uniformly and uses attention mechanism to calculate the relativity between the definition of relation type and the different relation instance.This aims to improve the importance of the relevant instance and weaken the irrelevant instance by weights.The experimental results show that the model based on the definition constraint and attention mechanism achieves better performance than the existing state-of-the-art model in the relation extraction task,which indicates that the definition constraint can effectively alleviate the noise problem in distant supervision and improve the performance of the relation extraction task.This paper introduces the relation type definition and uses the definition to restrict the noisy data generated by distant supervision method.The relation extraction system is optimized from the two aspects of sample filtering and multi-instance modeling.Both of these two methods have achieved certain effects.Among these two methods,the multi-instance model based on the attention mechanism is effective,and the F1 score is 2%higher than the baseline model on TAC-KBP task.At the same time,this paper also analyzes the advantages and disadvantages of the mainstream neural network framework in relation representation,and proposes a new neural network architecture for representing the relation examples.Experimental results prove that this method is superior to the state-of-the-art methods in relation extraction.