POP-CEE: Position-oriented Prompt-tuning Model for Causal Emotion Entailment

Zhihan Zhou, Xue Gu, Yujie Zhao, Hao XuACL 2024 · DOI

Introduction: The Challenge of Causal Emotion Entailment

In Natural Language Processing (NLP), recognizing what someone is feeling (Emotion Recognition in Conversation) is a well-established task. However, understanding why they feel that way—Causal Emotion Entailment (CEE)—presents a significantly harder challenge. Identifying the specific utterance that triggers an emotional response requires capturing both semantic meaning and the structural flow of dialogue.

Current approaches typically fine-tune models like BERT or RoBERTa. However, these methods suffer from two major gaps:

  1. The Task Gap: Standard fine-tuning often diverges from the "cloze" (fill-in-the-blank) tasks these models were originally pre-trained on.
  2. The Positional Gap: Different emotions exhibit distinct "causal signatures." For instance, happiness tends to be contagious across multiple turns, while surprise is almost always a reaction to the immediate preceding utterance. Existing models often use fixed-size windows that fail to account for these nuances.

Methodology: POP-CEE

To address these gaps, the authors propose POP-CEE (Position-oriented Prompt-tuning). This framework shifts from standard fine-tuning to a prompt-based approach that better aligns with the pre-training objectives of large language models.

The POP-CEE framework consists of an Emotion Prompt Module and a Position Constraint Module to predict causal links.

1. Emotion Prompt Module

Instead of a simple classification head, POP-CEE uses specialized templates to frame CEE as a cloze task. The model employs five distinct prompts:

  • Intra-Prompt: Targets cases where the cause is within the emotional utterance itself.
  • Happiness-Prompt: Captures the "contagious" nature of joy.
  • Surprise-Prompt: Focuses on immediate triggers.
  • Negative-Prompt: Designed for emotions like disgust or fear, which are rarely mirrored by the interlocutor.
  • Basic-Prompt: A generic template to ensure broad coverage and prevent overfitting.

2. Position Constraint Module

This module refines predictions by applying filters based on the statistical likelihood of a cause's location relative to the emotion. For example, it prioritizes immediate predecessors for surprise and speaker-internal history for negative emotions.

Key Insights: The Spatial Nature of Emotions

A core contribution of this work is the empirical observation that emotions have distinct "spatial" patterns within a conversation.

Statistical distribution of emotion causes relative to the target utterance (pos 0 is the utterance itself, pos 1 is the immediate predecessor).

As shown in the data, 84.6% of causes for "Surprise" occur exactly one utterance prior (position 1). Conversely, for "Disgust," the cause is found within the emotional utterance itself 83.9% of the time (position 0). By encoding these statistical priors into the Position Constraint Module, POP-CEE significantly improves its precision.

Experimental Results

The model was evaluated on the RECCON benchmark (DD and IE subsets). POP-CEE achieved state-of-the-art (SOTA) performance on both:

  • RECCON-DD: Achieved a Macro-F1 of 80.60%, outperforming complex models like the Window Transformer and standard RoBERTa-large fine-tuning.
  • RECCON-IE: Reached a Macro-F1 of 70.11%, a substantial improvement over previous baselines.

Ablation studies confirmed that the "Surprise-Prompt" and its associated positional constraints were the most critical components, as their removal led to the largest performance drops.

Conclusion

POP-CEE demonstrates that framing Causal Emotion Entailment as a position-aware prompt-tuning task is highly effective. By leveraging the latent knowledge in pre-trained models and accounting for the unique temporal signatures of different emotions, it achieves SOTA results without the need for heavy graph neural networks or complex attention mechanisms. Future work may involve automating prompt discovery and testing the framework on larger generative models.

Summary updated

BibTeX
@inproceedings{zhou2024pop,
  author = {Zhihan Zhou and Xue Gu and Yujie Zhao and Hao Xu},
  title = {{POP-CEE: Position-oriented Prompt-tuning Model for Causal Emotion Entailment}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-acl.248},
}
Figure