融合注意力机制和情感词位置的情感分析  

INTEGRATING ATTENTIONAL ENCODER AND POSITION OF SENTIMENT POLARITY FOR SENTIMENT ANALYSIS

作　　者：周晓雯 王晓晔[1] 孙嘉琪 于青[1] Zhou Xiaowen;Wang Xiaoye;Sun Jiaqi;Yu Qing(Tianjin Key Laboratory of Intelligence Computing and Novel Software Technology,Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津理工大学天津市智能计算及软件新技术重点实验室,天津300384

出　　处：《计算机应用与软件》2022年第12期180-186,194,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61872270,61572357)。

摘　　要：现有属性级情感分析方法主要是基于注意力模型,但都缺少情感词位置的提取,甚至忽略了情感目标在情感分类中的作用。针对上述问题,提出一种基于注意力编码机制和情感词位置的交互感知网络。通过预训练的BERT模型训练词向量;利用注意力编码机制对情感目标与上下文进行交互性学习,通过计算情感目标与全局文本的相关性来更新整体的权重分配。在情感词提取和分类中,通过扩展的情感词典对情感词进行提取,结合情感词和情感目标的位置信息生成<情感目标,情感词>,并通过Softmax层获得情感极性的预测。在AIChallenge2018的中文数据集上进行实验,该模型取得了比其他模型更好的分类效果。The existing aspect-level sentiment analysis methods are mainly based on the attention mechanism, but they are lack of extracting the position of sentiment words, and even ignore aspect terms in sentiment polarity classification. Aiming at these problems, this paper proposes an Interactive awareness network based on attentional encoder and position of sentiment polarity model. This paper used the pre-trained BERT model to train word embedding. The attentional encoder mechanism was proposed to learn the relationship between aspects and context interactively and updated the weight distribution by calculating the correlation between aspect terms and context. This model utilized extended sentiment lexicon to extract sentiment words and extracted the position information of sentiment words simultaneously. The prediction of emotional polarity was obtained through Softmax layer. Experiments on the AIChallenge2018 show that the model achieves better classification effect than other models.

关 键 词：属性级情感分析 注意力机制 情感位置感知 情感词典 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]