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.
Introduction: The Challenge of Emotional Transitions
In conversational AI, "empathy" requires more than just identifying a single emotion; it involves tracking the emotional journey of the speaker across multiple turns. Most existing models for empathetic response generation adopt a "coarse-grained" approach, concatenating previous dialogue turns into a single block of text to extract features.
The authors argue this approach is flawed because it ignores feature transitions. In a real conversation, a speaker's emotional state and focus can shift rapidly. For instance, a speaker might start by expressing excitement about a family visit but end by expressing embarrassment over a specific incident during dinner. A coarse-grained model often struggles with these shifts, leading to generic or emotionally inconsistent responses (e.g., "I'm glad you got her back" when the speaker is currently complaining about their boyfriend).
Methodology: Emp-RFT
Emp-RFT (Recognizing Feature Transitions) is built on the BART architecture and introduces three main innovations to capture these shifts:
- Feature Transition Recognizer (FTR): This component uses comparison functions (subtraction and multiplication) to measure how the current emotion, keywords, and utterance-level meanings differ from the previous two turns. This allows the model to "grasp" the flow of the dialogue.
- Fusing Context with Keywords: The model builds a dynamic keyword graph for each context. It uses a Graph Attention Network (GAT) to update node representations, ensuring that keywords related to the current situation receive more attention during response generation.
- Contrastive PPLM (Plug and Play Language Model): To avoid safe, generic responses (e.g., "I'm sorry to hear that"), a contrastive discriminator forces the model to actively incorporate specific predicted keywords, ensuring the response is tied to the actual topic.
Experimental Setup and Datasets
The researchers utilized the EmpatheticDialogues dataset, containing approximately 25,000 multi-turn dialogues grounded in 32 emotion labels. Since the original dataset lacked fine-grained labels for every turn, the authors performed Feature Annotation:
- Speaker Utterances: Emotions and "emotion cause" words were detected using a secondary model (EMO-CAUSE).
- Listener Utterances: Reactions (e.g., interpreting, exploring) were labeled using RoBERTa and communication mechanism models.
Key Results and Insights
Emp-RFT was benchmarked against state-of-the-art baselines, including MoEL, EmpDG, and Blender-based models.
- Automatic Evaluation: Emp-RFT outperformed all baselines in diversity metrics (Dist-1/2) and BERTscore, indicating more unique and semantically accurate responses.
- Human Evaluation: In A/B testing, human judges preferred Emp-RFT significantly. Against the MoEL model, Emp-RFT was preferred 74.4% of the time overall, rising to 82.2% for multi-turn instances.
- Multi-turn Robustness: Unlike many models that degrade as dialogues lengthen, Emp-RFT remained consistent, proving its transition-tracking mechanism successfully manages long-term dialogue flow.
Conclusion and Future Work
Emp-RFT demonstrates that the change in state between turns is often more informative for an empathetic listener than the content of the final turn alone. By treating keywords as anchors in a graph and using a contrastive approach to avoid genericness, the model generates more personal and relevant responses.
Future work includes reducing dependency on complex pre-trained feature extraction pipelines and simplifying the annotation process. The authors also suggest applying this transition-aware approach to complex open-domain social "chitchat" where emotional shifts are frequent.
Summary updated