基于对比主题模型和标签指导的对话情感识别

Dialogue Emotional Recognition Based on Comparative Topic Model and Label Guidance

在对话系统中,对话情感识别旨在预测对话中每个语句的情感标签,这个任务受多种因素的影响,如对话的主题、情感标签信息等。为了解决上述问题,提出了一个新的主题标签感知的图网络(TLGN)框架。该模型首先利用主题模型提取与情感相关的语义主题分布和学习标签指导的文本表示。然后,融合两种不同粒度的表示,作为文本的语义输入,输入到分类模型中,进行对话语句的情感预测。最后,在四个公开数据集上的实验结果表明,该模型的方法优于基准方法。

In a dialogue system,dialogue emotion recognition aims to predict the emotional labels for each statement in a conversation.This task is influenced by various factors such as the topic of the dialogue,emotional label information and so on.To address this issue,a new framework called Topic-aware Label Graph Network(TLGN)has been proposed.This model uses topic model to extract semantic topic distributions related to emotions and learns label-guided text representations firstly.Then,it fuses two representations of different granularity,used as semantic input to the text and input into the classification model for emotional prediction of dialogue statements.Finally,experimental results on four publicly available datasets show that the model method outperforms the benchmark methods.