基于双向注意力机制的问答情感分类方法

Sentiment Classification Towards Question-Answering Based on Bidirectional Attention Mechanism

情感分类是自然语言处理研究中的一项基本任务,旨在判别文本的情感极性。目前,情感分类相关研究主要针对句子、篇章和微博等文本形式。与以往研究不同的是,文中面向新颖的问答型评论展开情感分类。首先,收集并标注了大规模、高质量的问答型评论语料集;针对问答型评论的特点,提出了一种基于双向注意力机制的神经网络方法。具体而言,该方法首先通过双向LSTM对问题文本和答案文本分别编码,再通过双向注意力机制同时计算问题文本和答案文本的情感权重,最后通过情感权重计算得到问答型评论的情感匹配信息。实验结果表明,提出的方法在问答情感分类任务上达到了75.5%的准确率和61.4%的F1值,相较于其他基准方法有明显的提升。

Sentiment classification is a fundamental task in natural language processing,which aims at inferring the sentiment polarity of a given text.Previous studies for sentiment classification,mainly focus on sentence,document and tweet text styles.Different from these researches,this paper focused on a novel text style,i.e.,question-answering(QA)review,for sentiment classification.Firstly,a large-scale and high-quality QA review corpus was collected and built.Then,a bidirectional attention neural network for QA sentiment classification was proposed.Specifically,the question and answer text with Bi-LSTM were encoded respectively.After that,sentiment weights in question and answer text were calculated synchronously by employing bidirectional attention mechanism.Finally,the sentiment matching representation for each QA review with sentiment weights can be obtained.Empirical studies show that the proposed approach achieves a great result(75.5%in Accuracy and 61.4%in Macro F1),and has remarkable improvement compared with other baselines.