Modeling Intention, Emotion and External World in Dialogue Systems

Wei Peng, Yue Hu, Luxi Xing, Yuqiang Xie, Xingsheng Zhang, Yajing Sun

Abstract

Intention, emotion and action are important elements in human activities. Modeling the interaction process between individuals by analyzing the relationships between these elements is a challenging task. However, previous work mainly focused on modeling intention and emotion independently, and neglected of exploring the mutual relationships between intention and emotion. In this paper, we propose a RelAtion Interaction Network (RAIN), consisting of Intention Relation Module and Emotion Relation Module, to jointly model mutual relationships and explicitly integrate historical intention information. The experiments on the dataset show that our model can take full advantage of the intention, emotion and action between individuals and achieve a remarkable improvement over BERT-style baselines. Qualitative analysis verifies the importance of the mutual interaction between the intention and emotion.

Introduction

In human conversation, emotional states are rarely isolated occurrences; they are deeply intertwined with underlying goals and external responses. For example, if a speaker requests a glass of water (intention) and the listener provides it (external action), the speaker typically feels satisfied (emotion). Conversely, if the request is ignored, the speaker may feel frustration or anger. This causal link—where emotion is a psychological byproduct of how the external world reacts to internal intentions—is a fundamental aspect of human communication.

Conceptual framework of the interaction between intention, emotion, and external actions in a dialogue.

Despite this, most existing dialogue systems treat Intention Recognition and Emotion Prediction as separate classification tasks. While they may share text features, they often ignore the cumulative nature of "historical intention"—the concept that a speaker's previous goals continue to influence their emotional reaction to current responses. The proposed RelAtion Interaction Network (RAIN) addresses these gaps by explicitly modeling the interaction between these elements over time.

Methodology: The RelAtion Interaction Network (RAIN)

The RAIN framework mimics human psychological processes through two primary components: the Intention Relation Module (IRM) and the Emotion Relation Module (ERM). These modules work in tandem to process utterances and track the state of the conversation.

Architecture of the RelAtion Interaction Network (RAIN) featuring the Intention and Emotion Relation Modules.

Intention Relation Module (IRM)

The IRM utilizes a RoBERTa encoder to generate neural representations of the current utterance. A key innovation is the integration of an Intention Dictionary, a symbolic knowledge base of keywords (e.g., "would like," "can you") mapped to probability distributions of specific intentions. By combining neural embeddings with this interpretable symbolic signal, the model achieves higher accuracy in identifying the speaker's goal.

Emotion Relation Module (ERM)

The ERM recognizes that emotion depends on both the "External World" (the listener's actions) and the speaker's "Internal Intention." To capture this, the module employs an LSTM to track a sequence of historical intentions. An Intention Fusion Mechanism then combines these historical features with the current utterance's representation. This allows the model to "understand" that a speaker's happiness is specifically tied to the satisfaction of a prior request.

Experimental Setup and Results

The model was evaluated on the DailyDialog dataset, a standard benchmark for multi-turn dialogue. The researchers expanded the original intention labels into seven categories (request, suggest, command, accept, reject, question, inform) and utilized six emotion labels (happy, neutral, sadness, anger, content, disgust). To support the IRM, a subset of 2,046 conversations was manually annotated to construct the Intention Dictionary.

Performance Metrics

RAIN demonstrated significant improvements over state-of-the-art BERT-style baselines:

  • Intention Recognition: Achieved an F1 score of 72.93, outperforming RoBERTa-large by 2.02%.
  • Emotion Prediction: Achieved an F1 score of 64.07, representing a substantial 4.94% improvement over RoBERTa-large.

Ablation studies confirmed the importance of historical context; adding "Historical Intention" data provided a +2.98 F1 boost to emotion prediction. Furthermore, the model showed high efficiency, reaching peak performance within 5-6 training epochs, suggesting that explicit relational modeling helps the network capture relevant features more effectively than standard text classification.

Conclusion and Future Work

The RAIN framework successfully demonstrates that modeling the interplay between intention and emotion leads to more accurate dialogue understanding. By treating emotion as a reaction to the satisfaction or frustration of goals, the system moves closer to human-like interaction. Future research will focus on evolving the static Intention Dictionary into dynamic, automatically generated knowledge graphs and incorporating individual differences—such as personality traits and social context—to create more personalized dialogue agents.

Summary updated

Abstract

Intention, emotion and action are important elements in human activities. Modeling the interaction process between individuals by analyzing the relationships between these elements is a challenging task. However, previous work mainly focused on modeling intention and emotion independently, and neglected of exploring the mutual relationships between intention and emotion. In this paper, we propose a RelAtion Interaction Network (RAIN), consisting of Intention Relation Module and Emotion Relation Module, to jointly model mutual relationships and explicitly integrate historical intention information. The experiments on the dataset show that our model can take full advantage of the intention, emotion and action between individuals and achieve a remarkable improvement over BERT-style baselines. Qualitative analysis verifies the importance of the mutual interaction between the intention and emotion.
BibTeX
@inproceedings{peng2022modeling,
  author = {Wei Peng and Yue Hu and Luxi Xing and Yuqiang Xie and Xingsheng Zhang and Yajing Sun},
  title = {{Modeling Intention, Emotion and External World in Dialogue Systems}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=upi5i9hqgtJ},
}
Figure