📄 REG: Retrieval via Emotion Similarity for Guiding Empathetic Dialogue Generation

Summary

Retrieval-augmented empathetic chatbots usually pick example dialogues by semantic closeness. They can therefore pull in exemplars whose emotional tone contradicts the user's situation, such as a gloomy rain story for a cheerful one, which pushes the generator toward unsuitable replies.

REG makes retrieval emotion-aware. Each context–response pair is tagged with four Emotion Attributes, binned into Low/Mid/High levels:

• Sentiment, scored with VADER
• Emotional Presence, Interpretation and Exploration, each predicted by a separate classifier

Training has two stages. First, an initial generator G1 learns to respond conditioned on the context plus these attributes. Second, a retriever is trained by distillation from G1, with two additions:

• Token-level Retrieval uses token-level scores of G1's outputs (e.g., BLEU) as reward signals, which reduces noise from the coarse labels.
• Retrieval Candidate Augmentation injects randomized candidates, which increases exemplar diversity and helps avoid local optima.

The retriever and a stronger generator G2 are then trained jointly. On EmpatheticDialogues, REG retrieves emotionally aligned exemplars much more often than earlier retrievers. Human judges also prefer its responses over strong baselines, including a fine-tuned Llama-3.1. This shows that emotional matching, not only topical matching, matters for empathetic generation.

Key points

• The four Emotion Attributes are Sentiment, Emotional Presence, Interpretation and Exploration. They are drawn from prior empathy research (Sharma et al.; Watson et al.) and fed to the generator as auxiliary embeddings.
• The retriever is optimized by minimizing the KL divergence between its retrieval-score distribution and a distribution derived from G1's token-level metrics.
• Exemplar alignment was checked on 3,000 samples. REG retrieved aligned exemplars in 59.5% of cases, compared with 38.0% for LEMPEx, 36.6% for a fixed retriever and 8.6% for plain RAG.
• In human pairwise evaluation, REG won on Empathy against Llama-3.1-8B (SFT) by 58.0% vs 32.5%. Against CASE it won 56.3% vs 31.0%. Annotator agreement (κ) was about 0.47–0.57.
• Baselines included Transformer, KEMP, CASE, EmpCRL, PECER, LEMPEx, Exemplary-empathy and Llama-3.1 (SFT). All experiments used the EmpatheticDialogues dataset.