Abstract: Abstract: Sentiment classification in psychological counseling aims to obtain the sentiment polarity of the inquirer’s utterance, which can provide support for establishing psychological counseling AI assistants. Existing methods obtain the sentiment polarity of text through contextual information, failing to consider the sentiment transmission between the current sentence and the forward neighbor sentences in the dialogue record. To address the issue, this paper proposes a model based on Attention Over Attention mechanism. According to the model, historical sentiment words are assigned weights by temporal sequence. Thus it improved the accuracy of sentiment classification in psychological counseling. In a dialogue, historical sentiment word sequences of both sides were extracted by constructing a sentiment lexicon of mental health. Subsequently, the current sentence and two sequences of historical sentiment word were input into the Bidirectional Long Short-Term Memory (BiLSTM) to get corresponding feature vectors. The Ebbinghaus forgetting curve was used to allocate internal weights for the sequences of historical sentiment word. Both inertia features and interaction features were captured by Attention Over Attention (AOA) mechanism. Then, the above two features were input into the classification layer among with the text features, the sentiment polarity was carried out at final. Experimental results on public dataset Emotional First Aid Dataset show that the model proposed in this paper improves 1.55% compared with Caps-DGCN model in F1. Hence it can effectively improve the sentiment classification performance of psychological counseling.

Key words: psychological counseling, sentiment lexicon of mental health, the Ebbinghaus forgetting curve, Attention Over Attention mechanism, Bidirectional Long Short-Term Memory

摘要： 摘 要: 心理咨询场景下的情感分类旨在获得咨询者话语的情感倾向，为建立心理咨询AI助手提供支持。现有的方法利用语境信息获取文本情感倾向，但未考虑对话记录中当前句与前向近邻句之间的情感传递。针对这一问题，提出一种基于交互注意力机制的心理咨询文本情感分类模型，根据时序对历史情感词进行权重分配，进而提高分类准确率。利用构建的心理健康情感词典分别提取对话双方的历史情感词序列，再将当前句和历史情感词序列输入到双向长短期记忆网络（BiLSTM）获取对应的特征向量，并利用艾宾浩斯遗忘曲线对历史情感词序列进行权重分配。通过交互注意力机制（AOA）获得惯性特征和交互特征，并结合文本特征输入到分类层计算情感倾向。在公开数据集Emotional First Aid Dataset上的实验结果表明，相较于Caps-DGCN模型F1值提高了1.55%。可见，所提模型可以有效提升心理咨询文本的情感分类效果。

关键词: 心理咨询, 心理健康情感词典, 艾宾浩斯遗忘曲线, 交互注意力机制, 双向长短期记忆网络