How Do Bots Keep the Right Emotion in a Long Conversation?

A research team has devised a new way to make chatbots talk more consistently and emotionally. The key is the Verifiable Emotion Reward (VER), a reinforcement signal that uses pre‑trained emotion classifiers to score each turn of a conversation. In practice, the model gets a "point" whenever it keeps the emotion prescribed for its character, preventing random mood swings.

To test VER, scientists built CHARCO, a massive dataset of over 230,000 multi‑turn dialogues. Each exchange includes persona profiles, semantic contexts, and ten emotion labels, allowing the model to learn how a real role‑player would respond. After fine‑tuning large language models on CHARCO, the results were striking: bots displayed higher emotional consistency, better role fidelity, and smoother conversations.

Evaluations used lexical diversity metrics, automatic scoring with GPT‑4, and human assessments. All agreed that combining a purpose‑built dataset with the VER objective marks a significant leap for persona‑aligned chatbots, paving the way for more empathetic and believable virtual assistants.

The Verifiable Emotion Reward (VER) relies on frozen emotion classifiers, typically neural networks trained on labeled affective corpora (e.g., FER‑2013 or EmoBank). These classifiers output a probability distribution over a predefined set of emotions (joy, sadness, anger, surprise, disgust, fear, etc.). During training, the language model generates a response; VER computes the divergence (e.g., KL‑divergence) between the classifier's distribution and the target emotion associated with the character for that turn. The reward signal is then combined with the traditional language modeling loss, creating a hybrid objective that penalizes both syntactic errors and emotional deviations.

The CHARCO dataset was built via a crowdsourcing and automatic filtering pipeline. Participants received a persona sheet (age, hobbies, background) and a narrative context, then wrote dialogues of at least 10 turns, annotating the dominant emotion for each reply. Annotations were verified by cognitive psychology experts to ensure inter‑annotator consistency. The result is a balanced corpus with an almost uniform distribution of the ten emotions, enabling the model to learn emotional patterns even for less frequent states.

Evaluation metrics include: (i) Perplexity for linguistic fluency; (ii) Emotion Consistency Score (ECS), calculated as the average cosine similarity between target emotional embeddings and those generated; (iii) Role Fidelity Index (RFI), measuring the match between responses and persona traits. Additionally, validation studies with GPT‑4 assessed emotional coherence and credibility on a 5‑point Likert scale. Results show an average improvement of 0.35 points over baselines without VER, confirming the reward's effectiveness in mitigating emotional drift in long conversations.