Chinese Implicit Sentiment Analysis Based On Multi-Feature Fusion

Sentiment analysis is one of the most important applications in natural language processing.In the task of emotion analysis,it can be divided into explicit emotion analysis and implicit emotion analysis according to whether the expression level contains explicit emotion words.Explicit emotion analysis is the basis of emotion analysis and has been widely studied and applied,while implicit emotion analysis is a new research direction that has attracted much attention in recent years.Implicit emotion sentences do not contain obvious emotion words.Their wording is relatively objective and neutral,and the emotion expression is more implicit and euphemistic.Therefore,the traditional emotion analysis model is difficult to effectively identify implicit emotion.Considering the difficulties and characteristics of implicit sentiment analysis tasks,this paper mainly focuses on the following work:(1)A pre-training language model,Senti-Bert,which focuses on emotion analysis,is proposed.The semi-supervised method is used to conduct secondary pre-training on the selfproduced emotion data corpus.Compared with native Bert,Senti-Bert optimizes the training task and model structure.The Next Sentence Prediction task is changed to an emotion category task to improve the underlying information extraction ability and dataset utilization.At the same time,the model parameters are less,and the scale is smaller,which solves the problem of mismatch between the dataset and model scale.The experimental results show that the Senti-Bert model improves the effect by 1%-2% compared with the original Bert model.(2)A new implicit sentiment analysis model,ECISA-MFF,based on multi-feature fusion is proposed.Based on Sent-Bert,the multi-feature extraction method and crossproduct-based feature fusion strategy are further used in this model.According to the characteristics of implicit affective sentences,target sentences and contextual statements are distinguished.Sententi-Bert is used to represent the underlying word vector,and then different neural networks are used to extract multiple groups of different features.For the target sentence,the convolutional neural network is used to extract local features,and the threshold cycle unit is used to obtain time sequence features.For context statements,twoway gated loop units and attention mechanisms are used to extract the emotional characteristics of context content.Finally,the three extracted features are cross-productfused.The feature fusion strategy can fully integrate the multi-feature information and add the global fusion feature information while effectively preserving the local single feature,so as to solve the overfitting phenomenon caused by the direct splicing of multiple features.Experiments show that the ECISA-MFF model is effective in implicit sentiment analysis.(3)A multi-task and multi-feature implicit sentiment analysis model,CISA-MTMF,is proposed,which builds upon the ECISA-MFF model by adding a task to identify whether a sentence contains implicit sentiment.This approach not only effectively prevents overfitting using multi-task learning but also improves the efficiency of creating implicit sentiment analysis datasets.On the open dataset SMP2019-ECISA implicit sentiment analysis dataset,the model proposed in this study achieved an F1 value performance of 0.874,significantly higher than the reference model in the literature.It is also higher than the optimal model effect in SMP2021-ECISA task,which proves its effectiveness in the implicit sentiment analysis task.