Emp-RFT: Empathetic Response Generation via Recognizing Feature Transitions between Utterances

Wongyu Kim, Youbin Ahn, Donghyun Kim, Kyong-Ho Lee

Introduction

Empathetic response generation—the ability for an AI to understand a user's situation and respond with appropriate emotion—is a cornerstone of human-centric Natural Language Processing (NLP). However, existing models often struggle with multi-turn dialogues. Most current approaches extract features at a "coarse-grained" level, essentially concatenating the entire conversation history into a single block of text. This leads to a "dilution" of context where the model may lose track of how the speaker's mood and focus evolve over time.

For example, if a speaker starts a conversation excited about a visit but ends it feeling embarrassed by a specific event, a coarse-grained model might remain anchored to the initial "excited" state. Humans instinctively recognize feature transitions—changes in emotion, keywords, and meaning between utterances—to understand dialogue flow. The Emp-RFT framework is designed to bring this human-like recognition to AI.

Figure 1: An example of a multi-turn dialogue where the speaker's emotion shifts from excitement to embarrassment, illustrating the importance of tracking feature transitions for coherent responses.

Methodology

The authors introduce Emp-RFT (Empathetic response generation via Recognizing Feature Transitions), a framework that explicitly tracks how emotions and keywords shift from one turn to the next. The architecture is built upon a hierarchical structure using BART-base as the backbone for both word-level and utterance-level encoding.

Figure 2: The Emp-RFT architecture, detailing the Context Encoding phase with the Feature Transition Recognizer (FTR) and the subsequent Response Generation strategy.

Feature Transition Recognizer (FTR)

Instead of just encoding the current utterance, the FTR module uses comparison functions (subtraction and Hadamard product) to measure the difference between the current utterance's features—including emotion, keywords, and latent meaning—and those of the previous two turns. This allows the model to "grasp" the direction and momentum of the conversation.

Response Generation Strategy

To ensure responses are relevant and specific, Emp-RFT employs several advanced techniques:

  1. Keyword Fusing: The model builds a dynamic keyword graph and uses a Graph Attention Network (GAT) to emphasize specific terms within the context.
  2. Next Emotion/Keyword Detection: The model explicitly predicts which emotion and keywords should appear in the upcoming response.
  3. Contrastive PPLM: Inspired by "Plug and Play Language Models," the authors use a contrastive loss to ensure the generator actively incorporates the predicted keywords into the final response. This prevents the "generic response" trap, where models default to safe but hollow phrases like "I'm sorry to hear that."

Datasets and Evaluation

The model was trained and tested on the EmpatheticDialogues benchmark, which contains approximately 25,000 multi-turn dialogues grounded in 32 emotion labels. Because the original dataset lacks fine-grained per-utterance labels, the authors utilized pre-trained models (EMOCAUSE and RoBERTa) to annotate emotions and keywords for every turn. These automated annotations were validated by human workers, achieving a reliable Fleiss’ kappa of 0.55.

Key Results

Emp-RFT was compared against several state-of-the-art baselines, including MoEL, EmpDG, MIME, and a focused version of Blender.

  • Quantitative Performance: Emp-RFT achieved the best results in Diversity (Dist-1/2) and Semantic Similarity (F_BERT). While its Perplexity (PPL) was slightly higher than the massive Blender baseline, it outperformed all other models in its class.
  • Multi-turn Superiority: The performance gap between Emp-RFT and its competitors widened as dialogues grew longer. On multi-turn instances specifically, Emp-RFT showed a marked improvement in human ratings for Empathy and Relevance.
  • Human Preference: In A/B testing, human evaluators preferred Emp-RFT's responses over the strongest baseline (Blender+Focused S1) in 38.6% of multi-turn cases, compared to only 22.5% for the baseline.

Conclusion and Future Work

Emp-RFT demonstrates that the change in state between utterances is often more informative than the state itself. By calculating the delta between turns, the model identifies which parts of the dialogue history require the most attention.

While the model shows significant promise, it currently relies on automated labeling for its initial training data and introduces some computational overhead due to the GAT and contrastive PPLM components. Future research may focus on integrating external knowledge bases to provide deeper, "common-sense" empathy beyond the immediate keywords present in the text.

Summary updated

Abstract

Each utterance in multi-turn empathetic dialogues has features such as emotion, keywords, and utterance-level meaning. Feature transitions between utterances occur naturally. However, existing approaches fail to perceive the transitions because they extract features for the context at the coarse-grained level. To solve the above issue, we propose a novel approach of recognizing feature transitions between utterances, which helps understand the dialogue flow and better grasp the features of utterance that needs attention. Also, we introduce a response generation strategy to help focus on emotion and keywords related to appropriate features when generating responses. Experimental results show that our approach outperforms baselines and especially, achieves significant improvements on multi-turn dialogues.
BibTeX
@inproceedings{kim2022emp,
  author = {Wongyu Kim and Youbin Ahn and Donghyun Kim and Kyong-Ho Lee},
  title = {{Emp-RFT: Empathetic Response Generation via Recognizing Feature Transitions between Utterances}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.naacl-main.303},
}
Figure