Abstract:

To address the problems that many existing studies ignore the correlation between interlocutors’ emotions and sentiments， a sentiment boosting model for emotion recognition in conversation text was proposed， namely Sentiment Boosting Graph Neural network （SBGN）. Firstly， themes and dialogue intent were integrated into the text， and the reconstructed text features were extracted by fine-tuning the pre-trained language model. Secondly， a symmetric learning structure for emotion analysis was given， with the reconstructed features fed into a Graph Neural Network （GNN） emotion analysis model and a Bi-directional Long Short-Term Memory （Bi-LSTM） sentiment classification model. Finally， by fusing emotion analysis and sentiment classification models， a new loss function was constructed with sentiment classification loss function as a penalty， and the optimal penalty factor was adjusted and obtained by learning. Experimental results on public dataset DailyDialog show that SBGN model improves 16.62 percentage points compared with Dialogue Graph Convolutional Network （DialogueGCN） model， and improves 14.81 percentage points compared with the state-of-art model Directed Acyclic Graph-Emotion Recognition from Conversation （DAG-ERC） in micro-average F1. It can be seen that SBGN model can effectively improve the performance of emotion analysis in dialogue system.

Key words: Emotion Recognition in Conversations (ERC), sentiment classification, theme induction, Graph Neural Network (GNN), Bi-directional Long Short-Term Memory (Bi-LSTM)

摘要：

针对现有的许多研究忽略了说话人的情绪和情感的相关性的问题，提出一种情感增强的图网络对话文本情绪识别模型——SBGN。首先，将主题和对话意图融入文本，并微调预训练语言模型RoBERTa以提取重构的文本特征；其次，给出情绪分析的对称学习结构，将重构特征分别输入图神经网络（GNN）情绪分析模型和双向长短时记忆（Bi-LSTM）情感分类模型；最后，融合情绪分析和情感分类模型，将情感分类的损失函数作为惩罚以构建新的损失函数，并通过学习调节得到最优的惩罚因子。在公开数据集DailyDialog上的实验结果表明，相较于DialogueGCN模型与目前最先进的DAG-ERC模型，SBGN模型的微平均F1分别提高16.62与14.81个百分点。可见，SBGN模型能有效提高对话系统情绪分析的性能。

关键词: 对话情绪识别, 情感分类, 主题诱导, 图神经网络, 双向长短时记忆