基于情绪特定词向量的情绪分类算法

Emotion Classification Algorithm Based on Emotion-specific Word Embedding

情绪分析是自然语言处理领域的一个研究热点,其通过分析人们发布的文本推测人们的主观感受。情绪分类是情绪分析中的一个基本任务,旨在判断一个文本的情绪类别。对情绪分类来说,词语的表示具有决定性的作用。许多现有的词向量学习算法只对词语的上下文语义信息进行建模,而忽略了词语的情绪信息,这样会导致上下文相似但情绪相反的词语有相似的词向量。为了解决该问题,通过构建一个由两个基本网络(即文档-词网络和情绪图标-词网络)组成的异构网络来学习情绪特定的词向量。最后,在标注样本上训练一个LSTM分类器。实验结果表明了所提情绪特定词向量学习算法的有效性。

Emotion analysis is a hot research issue in the field of NLP,and it infersthe feelings of individuals through analyzing the text they have published.Emotion classification is a fundamental task in emotion analysis,which aims to determine the emotion categories in a piece of text.The representation of words is a critical prerequisite for emotion classification.Many intuitive choices of learning word embedding are available,but these word embedding algorithms typically model the syntactic context of words but ignore the emotion information relevant to words.As a result,words with opposite emotion but similar syntactic context tend to be represented as close vectors.To address the problem,this paper proposed a a heterogeneous network composed of two basic networks,i.e.,document-word network and emoticon-word network to learn emotion-specific word embedding.Finally,an LSTM network was trained on the labeled data.Empirical studies demonstrate the effectiveness of the proposed approach to learn emotion-specific word embedding.