An Entity Relation Classification Model Based On Self-Matching-Attention Mechanism

The entity in NLP means the recognizable nominal in human natural language,i.e.the name of substance,the name of person,the name of locations and so on.Classifying the relations between entities is a task to classify the semantic relations between entities in the sentences.It is the basis of the semantic recognition task in NLP and hold an important status.However,the usage of language is ever-changing,which brings a lot of difficulties to our task.Sometimes different pair of entities may share same semantic relation,while sometimes same pair of entities in different contexts may have different semantic relations.What's more,the various syntactic,such as the active and passive form of a sentence,also make it hard to achieve our task.In previous studies,we need to do preprocess the input sentence based on external knowledge before we send it into the classification system,such as extracting the dependency tree of a sentence to get the dependency path of a pair of entities,or tagging the words in the sentence with the Part-Of-Speech recognition tools.The results of these classification models exclusively rely on the performance of the preprocessing system.If the preprocessing system do not work effectively,it will bring a lot of noise to the classification system and make the system unstable.In this paper,we propose a new method based on the Multi-Self-Matching-Attention mechanism,it will help system enhance and focus on the words which are important to our classification task.This method significantly reduce the number of parameters of the model and lower the possibility of model overfitting.Experiment result shows that our method achieves an Macro-F1 Score of 85.34 on the data set of SemEval task8.The main research work in this paper includes following parts:1.Study and implement the method based on the traditional lexical features.Analyze the advantages and disadvantages of them and set the best score of these kind of methods as a baseline.2.Study and implement the neural network model based on syntax parse tree.Analyze the advantages and disadvantages of them and set the best score of these kind of methods as a baseline.3.Study and implement the method based on Attention Mechanism.Analyze the advantages and disadvantages of them and set the best score of these kind of methods as a baseline.4.Propose a new method base on Multi-Self-Matching-Attention mechanism to classify the sentiment between entities.Compare with other models,and analyze the advantages of our method.