Research On Relation Classification Method By Fusing Mutiple Features

Relation classification is an important task in the field of natural language processing,and it also provides technical support for the construction of knowledge graphs,question answering systems,and information retrieval.The emergence of deep learning technology has expanded the relevant research on relationship classification tasks,and further breakthroughs have been made in the experimental results.The relation classification method based on deep learning has gradually become one of the key research methods in this research.In recent years,researches on relationship classification based on deep learning mainly focus on the optimization of attention mechanism and the optimization of semantic information in task improvement.However,such methods still have the following problems:(1)many scholars in the study of relationship between classification using a complete sentence characteristics of complementary relations classification task,but not completely equivalent to the relationship between classification task statements classification task,context information related to the entity would affect the classification of between entities,which features some of the invalid relationship between the semantic information will become the noise of the classification performance;(2)When the relational classification model is trained on a data set with a small corpus,the model cannot learn a more complete semantic representation due to the small amount of information in the data set itself,and the extraction source of semantic information needs to be further expanded;(3)The experimental performance of deep learning models is relatively dependent on the features extracted from the model.When the features extracted from the model are relatively single,the performance of the relational classification model will be limited to further improvement.To solve the above problems,this thesis proposes a relationship classification method that integrates multi-feature information,and improves the relationship classification method from the aspects of deep learning model and semantic features.In this thesis,SEBN model is constructed on the basis of BERT model and Nystr(?)mformer attention mechanism,and the SEBN model is used to focus on the semantic information in the corpus that is more conducive to relational classification.The Type-SEBN model is used to learn many kinds of semantic information,such as semantic information,entity information and entity Type information,and the validity of the semantic information and entity Type information proposed in this thesis as well as the validity of the semantic information fusion method is proved through relevant experiments.The relevant work of this thesis is as follows:(1)By referring to the industrial-grade entity type classification standards such as spa Cy and Stanford Core NLP,the classification standards were selectively optimized according to the research needs,and an entity type classification standard,imp Spa Cy,was constructed to meet the needs of this study.The specific meaning of the corresponding entity in the sentence is clarified in order to label the entity type of each primitive sentence in the Semeval-2010 Task 8 dataset in the form of manual annotation by referring to the imp Spa Cy standard.A class vector generation algorithm is proposed,which extracts the class vector of entities from the data set and introduces it into the model as an external corpus for learning,making it an important semantic feature of important relationship classification supplemented by the completion of experiments.(2)SEBN model is constructed by combining BERT model and Nystr(?)mformer attention mechanism.Sentences in Semeval-2010 Task 8 corpus are encoded into word vectors through BERT model,and then all generated word vectors are input into Nystr(?)mformer layer together.The Nystr(?)mformer attention mechanism is used to focus on important semantic information that is beneficial to relation classification.In order to improve the effect of subsequent tasks,the effect of semantic information extraction of BERT model is improved.(3)The SEBN model is used to learn the compound semantic information which is formed by the fusion of sentence meaning information,entity information and entity Type information,and the final model Type-SEBN of this thesis is constructed.After combining,the semantic features,entity features and entity type features are respectively sent to the full connection layer and the softmax layer for processing,and the relationship types between the specified entity pairs are obtained.(4)Through setting experiment and analysis,it is proved that the model in this thesis is superior to the reference model in relation classification task,and it is also proved that the compound semantic feature proposed in this thesis combines the semantic feature,entity feature and entity type feature.