| Sentiment analysis in natural language processing(also known as opinion mining)aims to automatically discover the sentiment contained in texts,and is widely used in fields such as public opinion analysis,sentiment dialogue,and commodity service review analysis.Different from document-level and sentence-level sentiment analysis,aspect-level sentiment analysis needs to mine the corresponding sentiment color at the entity or aspect level,and can be divided into single-task and multi-task aspect-level sentiment analysis according to whether the aspect and sentiment are jointly extracted.The former is not suitable for practical applications because it can only solve a single problem;the latter faces problems such as insufficient information interaction between subtasks and unsupervised attention mechanism.This paper mainly studies the problems existing in multi-task aspectlevel sentiment analysis,including the application of emotional referential meaning to the joint extraction of aspect sentiment,the coding mechanism of feature fusion between sub-tasks,and the training algorithm of counterfactual attention mechanism.The results obtained are as follows:(Ⅰ)Propose an aspect sentiment interactive fusion model(Aspect Sentiment Interactive Fusion,ASIF).Aiming at the problem that sub-tasks cannot fully integrate coding information in the existing joint task model of sentiment analysis,ASIF combines the referential meaning of emotional words in linguistics to design a feature interaction fusion mechanism between sub-tasks to promote the deep fusion of sub-tasks with each other’s feature encoding.The model adopts the pre-training model as the shared layer of the two subtasks,and finally completes the respective outputs after fusing the features of the subtasks:aspect category recognition and sentiment classification.This paper conducts several sets of experiments to verify the effectiveness of the interactive fusion mechanism.Comparing with other top-level models in other fields,the performance of the model proposed in this paper is superior to the above models on both sentiment analysis datasets,which proves the effectiveness of the method in this paper.(Ⅱ)Propose a counterfactual attention mechanism training algorithm.In view of the unsupervised characteristics of the attention mechanism in the existing model,combined with the counterfactual attention mechanism,adding supervision through attention annotation,and designing a counterfactual attention mechanism training algorithm to optimize the performance of the model.Considering the overall effect and labeling overhead,the sentiment dictionary is used to label the important words related to sentiment analysis in the two datasets,and the counterfactual attention mechanism is applied to the fusion layer of the model.On the basis of the test,the optimal level of applying the counterfactual mechanism is obtained.Through experiments on two datasets,it is proved that this method can optimize the performance of the model.(Ⅲ)Design and implement a sentiment analysis system for joint extraction of aspect sentiment.The system is implemented based on Python and Java languages,including model loading module,preprocessing module and visualization module.The model trained on the Semeval 2014 Restaurant data set was applied to the system,and the automatic test script was written and the Restaurant data set was used for experimental verification.The experimental results show that the sentiment analysis system designed in this paper has certain usability and good performance. |
