With the continuous strengthening of supervision, people’s way of expressing sentiment has gradually evolved from straightforward to obscure, and the task of sentiment analysis becomes more difficult. In view of the huge difficulties encountered by the pre-trained language model and fine-tuning stage model used for sentiment analysis when dealing with implicit sentiment text, this paper proposes a method ImplicitPrompt based on prompt learning model（a model that bridges the pre-trained model and fine-tuning stage model） and category differential loss, which makes obscure expression straightforward, distinguishes the distance between different sentiment labels, and improves the accuracy of implicit sentiment classification. The paper conducts extensive experiments under both standard supervised settings and low-resource settings in English standard aspect-based sentiment analysis datasets. The empirical experimental results demonstrate the effectiveness of the proposed method.

用户评论往往同时包含多个方面、多种情感,如何正确判断一条评论中不同方面的情感倾向性是方面情感分析的难点.文中提出基于词嵌入与记忆网络的方面情感分类.通过在记忆网络的不同模块引入方面词向量,加强方面词的语义信息,指导注意力机制捕捉方面相关的上下文信息,提升方面情感分类效果.在SemEval 2014任务4的短文本英文评论数据集和文中标注的长文本中文新闻数据集上实验表明,文中方法分类效果较好,在记忆网络框架下引入方面词嵌入信息是有效的.