Do You Understand How I Feel?: Towards Verified Empathy in Therapy Chatbots

TL;DR

This work tackles the challenge of verifiable empathy in therapy chatbots by pairing NLP-derived dialogue features with a Stochastic Hybrid Automaton (SHA) model of dyadic therapy sessions. Empathy properties are specified and verified using MITL-based probabilistic checks via Statistical Model Checking in Uppaal SMC, with strategy synthesis to steer agent behavior toward empathetic outcomes; properties can be expressed as $P_M(\diamond_{\leq \tau}\ \psi)$ or $P_M(\square_{\leq \tau}\ V \in [V_{min}, V_{max}])$. Preliminary results show the SHA captures core therapy dynamics with fidelity and that a High Valence policy improves empathy-satisfaction probabilities over baselines. This framework enables traceable, verifiable empathetic therapy agents, potentially enabling safer deployment of mental-health chatbots in socio-critical settings.

Abstract

Conversational agents are increasingly used as support tools along mental therapeutic pathways with significant societal impacts. In particular, empathy is a key non-functional requirement in therapeutic contexts, yet current chatbot development practices provide no systematic means to specify or verify it. This paper envisions a framework integrating natural language processing and formal verification to deliver empathetic therapy chatbots. A Transformer-based model extracts dialogue features, which are then translated into a Stochastic Hybrid Automaton model of dyadic therapy sessions. Empathy-related properties can then be verified through Statistical Model Checking, while strategy synthesis provides guidance for shaping agent behavior. Preliminary results show that the formal model captures therapy dynamics with good fidelity and that ad-hoc strategies improve the probability of satisfying empathy requirements.

Figures (3)

• Figure 1: Envisioned approach: solid and dashed arrows represent training and inference tasks, respectively.
• Figure 2: Dialogue model.
• Figure 3: Transitions compliance of observations from test data with UPPAAL estimated confidence bounds