基于对话者语句交互图神经网络的对话情感分析

Conversational sentiment analysis based on speaker utterance interaction graph neural network

对话情感分析任务旨在通过理解人类在对话中表达情绪的方式,结合对话内容及对话者信息,识别对话中的每一个语句的情感分类。不同于其余文本情感分析任务,对话情感分析需要建模对话中的顺序上下文语境。然而,如何更直观有效地建模对话语境,并且充分考虑对话参与者的情绪变化,以提高对话情感分析任务的准确率等问题仍有待探索。因此,本文提出一种基于对话者语句交互图神经网络的对话情感分析模型。首先,通过微调RoBERTa预训练语言模型提取对话文本的语句特征和对话者信息特征;其次,使用Bi-GRU建模对话文本的序列上下文语境,获得上下文语句特征;最后,融合上下文语句特征和对话者信息特征构建对话者语句交互图神经网络模型。在公开数据集MELD上的实验结果表明,与其他基线模型相比,本文所提模型取得了更好的实验性能。

Conversational sentiment analysis aims to identify the sentiment classification of each utterance in a conversation by understanding the way human express emotions and combining the conversation and speaker information.Different from the other text sentiment analysis tasks,conversational sentiment analysis needs to model the sequential context.However,issues such as how to model the conversation context more intuitively and effectively,and fully consider the sentiment changes of participants to improve the accuracy of the sentiment analysis remain to be explored.Therefore,this paper proposes a conversational sentiment analysis model based on speaker utterance interaction graph neural networks(SUDG).Firstly,the utterance features and speaker information features are extracted by fine-tuning the RoBERTa pre-trained language model.Then,Bi-GRU is used to encode the sequential contextual to obtain the contextual utterance features.Finally,the fusion of the contextual utterance features and the speaker information features constructs the speaker utterance interaction graph neural network model.Experimental results on the public dataset MELD show that the proposed model achieves better experimental performance compared with other baseline models.