Research And Implementation Of Robust Multi-task Sentiment Analysis Technologies

The openness of the Internet allows users to comment on the activities they have participated in online,which can also provide decision-making reference for other users.The explosive growth of information makes it difficult to analyze artificially.Thus,it is important to make automatic sentiment analysis of these comments from users,both theoretically and practically.Different from document-level and sentence-level sentiment analysis,aspect-level sentiment analysis is a fine-grained sentiment analysis task.The basic purpose of the task is to find evaluation objects and the corresponding sentiment polarity,which makes it more challenging.Aspect-level sentiment analysis has many sub-tasks,among which the most basic sub-tasks are aspect word extraction and aspect sentiment classification.This thesis discusses how to build a robust aspect-based sentiment analysis model from three aspects: building a stronger model,defense measures against general word substitution attack and task related robustness.The details are as follows:Firstly,this thesis proposes a BERT-based end-to-end opinion mining method for tourism comments,which models the joint task as a unified sequential labeling task.The joint task was enhanced by Bert and span extraction method with replication ability.And it was applied to real Chinese tourism comments.Considering the problem of missing aspect words in Chinese comments,this thesis modifies the target of the joint task to get a more complete opinion expression.Compared with the classical sequential labeling methods,the method in this thesis has achieved better results.In addition,this thesis conducts a data set for comments on Chinese tourism website,which can be used for the research of various sub-tasks of aspect-level sentiment analysis in Chinese.Secondly,due to the vulnerability of deep neural network,this thesis proposes a defense method based on adversarial training and certified robustness technology to defend against general word replacement attacks.Specifically,the model is enhanced by adversarial training with the provable input boundary constraints which are constructed by the certified robustness technique based on randomized smoothing.Experiments on the aspect sentiment classification models show that the method is effective on word replacement attack.The internal structure of the model can be ignored,which means it can be successfully applied to the aspect sentiment classification task with complex model structure.In addition,the contribution of adversarial training and certified robustness technology is analyzed separately.Finally,although single sub-task has achieved good results,its robustness is still worth exploring.Aspect-level sentiment analysis has a specific task purpose,and the general word replacement attack is not enough specific for the task.To solve this problem,this thesis generates a robust problem for aspect-level sentiment analysis task,and proposes a sentiment obfuscation attack method to explore the corresponding robustness problem.Specifically,the attack method is to construct adversarial examples with replacing target opinion words with synonyms and flipping the sentiment polarities of non-target opinion words.In the experiments,we firstly evaluate the quality of the adversarial samples.And the attack method is applied to different models,which is proved to be effective,and reflects the robustness problem of the current advanced models.Finally,according to the performance of the model,the influence of different structures and fused information is discussed.