Research On Interpretable Sentiment Analysis Model Based On Deep Learning

The rapid development of social networks has provided a broad platform for people to express and share personal opinions,and more and more people have expressed opinions and expressed emotions on the Internet.Based on this,potential users can browse these subjective comments to understand public opinion on a certain thing or product.Therefore,how to use natural language processing(NLP)technology to analyze the sentiment tendency of short texts on the internet has become a hot spot for researchers.Currently,deep learning methods have produced state-of-the-art results in many sentiment analysis tasks.Most of these studies use convolutional neural networks(CNN)and recurrent neural networks(RNN).In particular,derivative networks of RNN such as LSTM((Long Short-Term Memory)and GRU(Gated Recurrent Unit)can not only better solve the problem of text sequences,but also avoid the vanishing gradient to a certain extent,thereby obtaining better classification results.However,deep learning models are usually used as a “black box”,that is,the model only gives the classification results,but does not make an understandable explanation of the model's classification results and decisions,which makes users unable to fully trust the model.As a result,it cannot help users make rational decisions.At the same time,research shows that most of the existing sentiment analysis models only focus on the relationship mining between context features,and fall short in effectively incorporating into the deep neural network external knowledge,e.g.,affective or commonsense knowledge,that could directly contribute to the decisions of the classifier.In view of the above problems,this paper proposes two interpretable sentiment analysis models based on deep learning: 1)Attention Networks for Aspect-level Sentiment Analysis(DANSA),for aspect-level sentiment analysis tasks,using self-attention and multi-head attention to simultaneously obtain the global structure information of the text and the information related to aspects,it is solved that CNN has difficulty obtaining global information and the training time-consuming of RNN is too long,and the degree dependence between words graduallydecreases as the distance increase.Moreover,the networks use multiple attention mechanisms to obtain the importance of each word to the decision of the classifier,so as to explain the classification results of the model.The extensive experiments on the SemEval2014 datasets and the Twitter datasets,compared with the current mainstream deep learning methods,DANSA has achieved better classification results while giving decision explanations through the attention mechanism.2)A Generative Fine-Graind Explanation Model incorporating External Commonsense Knowledge for Sentiment Analysis(GECKSA),for the document level sentiment analysis task,it uses the apriori algorithm and frequency distribution to generate a fine-grained explanation of the text,and obtains a sentiment score of the fine-grained explanation by fusing commonsense knowledge,and then uses the score to guide the decision process of the model classifier.Evaluating GECKSA on several neural network baseline methods for the Hotel Reviews dataset and the Amazon Food Reviews datasets,including the CNN-based model,the LSTM-based model and the Transformer-based model.Experimental results show that GECKSA can not only generate fine-grained explanation but also substantially improves the performance over baseline system.