Predicting Evoked Emotions in Conversations

TL;DR

This work defines Predicting Evoked Emotions in Conversations (PEC), a task to forecast the emotion on the next turn given prior dialogue, and introduces two neural architectures, BiLSTM-PEC and DGCN-PEC, to capture sequence information, self-dependency, and recency. Through comprehensive experiments on dyadic and group datasets, the study demonstrates that recency and self-dependency are crucial for accurate prediction, with DGCN-PEC excelling in text- and emotion-text-based sequences due to its speaker-aware graph modeling. The results offer practical guidance for empathetic dialogue systems and proactive toxicity detection, showing when graph-based vs. sequence-based models are preferable and highlighting the value of self-specified emotional history. The authors also emphasize reproducibility and ethical considerations, and plan to release code to support further research.

Abstract

Understanding and predicting the emotional trajectory in multi-party multi-turn conversations is of great significance. Such information can be used, for example, to generate empathetic response in human-machine interaction or to inform models of pre-emptive toxicity detection. In this work, we introduce the novel problem of Predicting Emotions in Conversations (PEC) for the next turn (n+1), given combinations of textual and/or emotion input up to turn n. We systematically approach the problem by modeling three dimensions inherently connected to evoked emotions in dialogues, including (i) sequence modeling, (ii) self-dependency modeling, and (iii) recency modeling. These modeling dimensions are then incorporated into two deep neural network architectures, a sequence model and a graph convolutional network model. The former is designed to capture the sequence of utterances in a dialogue, while the latter captures the sequence of utterances and the network formation of multi-party dialogues. We perform a comprehensive empirical evaluation of the various proposed models for addressing the PEC problem. The results indicate (i) the importance of the self-dependency and recency model dimensions for the prediction task, (ii) the quality of simpler sequence models in short dialogues, (iii) the importance of the graph neural models in improving the predictions in long dialogues.

Figures (9)

• Figure 1: Transition matrix showing the transition probabilities of one emotion to another. $e_{n+1}$ and $e_{n-1}$ come from the same user; $e_n$ pertains to another user.
• Figure 2: Overview of the proposed DGCN-PEC model architecture.
• Figure 3: Comparing the wSLB and wOLB sequence models in DAILYDIALOG using BiLSTM-PEC.
• Figure 4: Macro-avg F1 score using E-BiLSTM-PEC for the MELD group conversation dataset. Each color represents one of the six characters in the dataset.
• Figure 5: Emotion profiles of three example characters ( Chandler, Joey, Pheobe) coming from the MELD dataset.