| As an indispensable and fine-grained task in sentiment analysis,aspect-level sentiment classification aims at automatically predicting the sentiment polarity(e.g.positive,negative,neutral)of an input sentence at the aspect level.In aspect-level sentiment classification,it is prevalent to equip dominant neural models with attention mechanisms,for the sake of acquiring the importance of each context word on the given aspect.However,such a mechanism tends to pay attention to a small number of frequent words with sentiment polarities while many infrequent important words are ignored.On the other hand,due to the lack of corpus for aspect-level sentiment classification task,it also greatly increases the training difficulty of neural networks.To deal with these issues,this article proposes two models as follows.1.An aspect-level sentiment classification model with progressive self-supervised attention learning.Specifically,at each iteration,for each instance,the context word is identified with the greatest effect on the sentiment prediction according to its attention weights or gradient and then extract it as the one with active/misleading influence on the correct/incorrect prediction of the instance.To exploit them to refine aspect-level sentiment classification models,the conventional training objective is augmented with a regularization term that encourages the models to not only equally consider the extracted active context words but also decrease weights of those misleading ones.2.An aspect-level sentiment classification model based on BERT,which is a language model pre-trained on large-scale text corpus.The pre-training model BERT is able to help the classification model alleviate the problem of lack of corpus.Furthermore,a word saliency metric based on partial gradient is proposed to identify the most important context word for the BERT-based classification model.These two approaches have been compared with the widely used contrast models in the aspect-level sentiment classification task,and demonstrated the effectiveness of the proposed approaches through experimental analysis. |
