Research On Feature Extraction-oriented Transfer Learning Method In Emotional Analysis

For a long time,people have tried to extract useful knowledge from text.Information extraction is also s hot topics in natural language processing.Attributed to the improvement of computer artificial intelligence,it is now possible for us to extract more abstract information from text,like emotional information.At present,the analysis method based on deep neural network has achieved good results in text emotion analysis.However,since the training process of the model is based on the traditional machine learning,which assumes the test samples and training samples come from the same domain.It will lead to a poor performance of the model,causing by the data dependence problem,when using on another dataset set less related with its training domain.With the rapid increase of network data,training methods based on traditional assumptions are difficult to meet the needs of real-world applications.Therefore,we discuss the possible factors that affect the prediction effect of the crossdataset domain model,and compares the effect of char-based model and word-based model in Chinese emotional classification tasks.From the perspective of feature extraction,this thesis combines transfer learning and emotional analysis models to propose a deep transfer network with a domain separation adaptive method.Experiments on emotional classification tasks,in this thesis,prove the effectiveness of this scheme.The main research and contributions are as follows:1.Propose a dynamic word embedding scheme based on ALBERT.Combined with previous results,this thesis finds that using multi-layer dynamic word embedding to build text feature extraction model can solve the representation problem of unknown out-of-vocabulary words,and can also increase the interpretability of the model.In addition,the pre-training finetune of Albert can also increase the generalization of the model.The accuracy of ALBERT dynamic word embedding model achieves 93.52% in the multi-type emotion dataset,and the comparative experiment proves that this method can have better feature representation.2.Propose a deep transfer network model with domain separation adaptive method.The proposed model combines emotional analysis and transfer learning,which includes dynamic word embedding layer,feature extraction layer,domain adaptation layer,feature fusion layer and classification layer.Domain adaptation layer is divided into a public network and a private network,learning the public features and private features of the source domain and the target domain respectively by using an adaptive manner to narrow the distance between two domains.Experiments on Chinese and English datasets prove that the proposed model can realize the migration prediction between different data domains.