基于ATT-DGRU的文本方面级别情感分析

方面级别情感分类是针对给定文本、分析其在给定方面所表达出的情感极性。现有的主流解决方案中,基于注意力机制的循环神经网络模型忽略了关键词邻近上下文信息的重要性,而结合卷积神经网络(Convolutional Neural Network,CNN)的多层模型不擅长捕捉句子级别的长距离依赖信息。因此,提出了一种基于截断循环神经网络(Disconnected Gated Recurrent Units,DGRU)和注意力机制的方面级别情感分类网络模型(Attention-Disconnected Gated Recurrent Units,ATT-DGRU)。DGRU网络综合了循环神经网络和CNN的优点,既能捕捉文本的长距离依赖语义信息,又可以很好地抽取关键短语的语义信息。注意力机制在推断方面情感极性时捕获每一个单词与给定方面的关联程度,同时生成一个情感权重向量用于可视化。ATT-DGRU模型在中文酒店评论数据集上进行ACSA任务,任务结果表明,其二分类、三分类准确率分别达到91.53%,86.61%;在SemEval2014-Restaurant数据集进行ATSA任务,任务结果表明,其二分类、三分类准确率分别可达90.06%,77.21%。

Aspect-level Sentiment Analysis of Text Based on ATT-DGRU

Aspect-level sentiment classification is to analyze the sentiment polarity in a given aspect for a given text.In the exis-ting mainstream solutions,the attention mechanism-based cyclic neural network model ignores the importance of keyword pro-ximity context information,and the CNN multilayer model is not good at capturing sentence-level long-distance dependency information.This paper proposes an aspect-level emotion classification network model based on disconnected gated recurrent units(DGRU)and attention mechanism,which is abbreviated as ATT-DGRU.The DGRU network used in this model integrates the advantages of circular neural network and CNN.It can not only capture the long-distance dependent semantic information of text,but also extract the semantic information of key phrases.Attention mechanism is used to capture the importance of each word to a specific aspect when deducing the sentiment polarity of a specific aspect,meanwhile generates an emotional weight vector,which can be visualized.Accuracies of two-class and three-class of ATT-DGRU model constructed in this paper can reach 91.11%and 87.76%respectively in ACSA task on Chinese hotel review datasets.Accuracies of two-class and three class of ATT-DGRU model can reach 77.21%and 90.06%respectively in ATSA task on SemEval2014-Restaurant dataset.