Research On Unknown Target Position Detection Method Based On Domain Generalization And Information Enhancement

In the past few years,although the research on stance detection has made great progress,most of the existing work only focuses on the situation that the distribution of training set and test set is consistent,but ignores the feature of social platform topics,namely the fast updating speed.Therefore,dealing with new and unfamiliar topics is extremely important.For this reason,this paper applies deep neural network as the basic model to explore the method of stance detection task in dealing with unknown topics.Specifically,this paper will explore from the following three aspects:(1)Unseen target stance detection method based on Domain Generalization: This article will unknown target position detection task as a classification task,the first to use the two Bi-LSTM characteristics of topic and comment are extracted respectively,and encoding the use conditions to the subject of information fusion into text vector said,finally,introduced against neural network generalization method in the area,so that the model can capture information characteristics of a more general said.The experimental results show that considering the topic as a domain and introducing the domain generalization method can improve the detection performance of the model under unknown topics.(2)Unseen target stance detection method based on data augmentation.On the basis of the domain generalization method introduced above,this paper attempts to use the method of translation back to augment the training data.Firstly,the English data training set used in this paper is translated into other language,such as Chinese,using the existing API,and then translated back to English.In order to make the augmented data retain the previous label characteristics,this paper introduces the Cbert model,uses the random occlusion vocabulary and regenerates it to achieve data augmentation.The experimental results show that data augmentation benefits the task of this paper.(3)Unseen target stance detection method based on external knowledge: in order to make the system take advantage of some obscure connections between words,this paper tries to enrich information representation by using the pre-training model and adding explanatory information to the topic on the basis of the above,so as to further achieve the purpose of introducing external knowledge.The experimental data show that using the word vector generated by the pre-training model can improve the detection performance of the model with external knowledge.