Fusing Distinguish Degree Neural Networks For Relational Classification

Relational classification is an important natural language processing(NLP)task.It plays an important role in many natural language processing tasks such as information extraction,question and answer system,knowledge base construction,etc.Most of the past methods regard the relational classification task as a simple multi-classification task,without considering similar problems between categories and categories.For the two kinds of relationships with different semantic directions of the same semantic relationship,since the semantic relations are the same,their expressions are often similar,so these two kinds of relationships are easily confused.In order to solve the problem that the relationship categories with the same semantic relationship but different entity directions are easily confused,this paper proposes a neural network that combines discriminant information for relationship classification.In this model,the discriminant information is introduced to distinguish the relationship categories in which the semantic relationships are the same but the entity directions are different.We transform the spatial direction of the entity into a mathematical vector direction by subtracting the entity word vector,and the result of subtracting the entity word vector as the discrimination information.The model is divided into three modules:sentence representation module,relationship differentiation module and differentiation fusion module.In the sentence representation module,Bi-LSTM is used as the encoder to encode the semantic information of the sentence.We embed the word embedding with the location feature as the input to Bi-LSTM additional features.In the relationship distinguishing module,the result of subtracting the entity word vector is used as the input of the relationship distinguishing module,and the degree of discrimination is obtained after nonlinear transformation.The discriminative degree fusion module fuses the output of the sentence representation module with the output of the relationship distinguishing module.In this paper,two fusion methods are used for feature fusion.One is a cascade-based feature fusion method,which outputs a hidden layer state of the forward LSTM of e2 and a hidden layer state of the backward LSTM of el in the sentence representation module,and then cascades with the distinguishing feature of the relationship distinguishing module.The cascaded features are sent to the classifier for classification.The other is CNN-based feature fusion method.The other is based on the feature fusion method of convolutional neural network.The output of the sentence representation module is paralleled with the output of the relationship distinguishing module as the input feature of the convolutional neural network,and the maximum pooling and chunk pooling are used in the pooling layer.Finally.the experimental results of the model show that the discriminant features proposed in this paper are effective in distinguishing the categories that are easily confused.In addition,our proposed model yields an F1 value of 84.8%without any additional features and NLP analysis tools.The contributions of this article are as follows:(1)We have found that the relationship categories with the same semantic relations but different entity directions are easily confused,and propose the concept of discrimination and use it to solve this problem.(2)Adding the deformed max-margin function and the cross entropy function as a new loss function improves the performance of the model to some extent.(3)Compare the effects of two different fusion methods on relationship classification and analyze the reasons.(4)We conducted experiments on the standard dataset SemEval-2010 Task 8 of the relationship classification.The experimental results show that the discriminant feature proposed in this paper can effectively improve the performance of the relational classification model.Compared with the Bi-LSTM baseline model,the F1 value of the model is increased by 4.4%after adding the discriminant feature.The model proposed in this paper achieves an F1 value of 84.8%without any additional features and NLP tools.