Daniel Cahn

Ziyi Zhu, Olivier Tieleman, Caitlin A. Stamatis et al. · cambridge

Realistic user simulation is crucial for training and evaluating task-oriented dialogue (TOD) systems, yet creating simulators that accurately replicate human behavior remains challenging. A key property of effective simulators is their ability to expose failure modes of the systems they evaluate. We present an adversarial training framework that iteratively improves user simulator realism through a competitive dynamic between a generator (user simulator) and a discriminator. Applied to mental health support chatbots, our approach demonstrates that fine-tuned simulators dramatically outperform zero-shot base models at surfacing system issues, and adversarial training further enhances diversity, distributional alignment, and predictive validity. The resulting simulator achieves a strong correlation between simulated and real failure occurrence rates across diverse chatbot configurations while maintaining low distributional divergence of failure modes. Discriminator accuracy decreases drastically after three adversarial iterations, suggesting improved realism. These results provide evidence that adversarial training is a promising approach for creating realistic user simulators in mental health support TOD domains, enabling rapid, reliable, and cost-effective system evaluation before deployment.

Daniel Cahn

The Shout Crisis Text Line provides individuals undergoing mental health crises an opportunity to have an anonymous text message conversation with a trained Crisis Volunteer (CV). This project partners with Shout and its parent organisation, Mental Health Innovations, to explore the applications of Machine Learning in understanding Shout's conversations and improving its service. The overarching aim of this project is to develop a proof-of-concept model to demonstrate the potential of applying deep learning to crisis text messages. Specifically, this project aims to use deep learning to (1) predict an individual's risk of suicide or self-harm, (2) assess conversation success and CV skill using robust metrics, and (3) extrapolate demographic information from a texter survey to conversations where the texter did not complete the survey. To these ends, contributions to deep learning include a modified Transformer-over-BERT model; a framework for multitask learning to improve generalisation in the presence of sparse labels; and a mathematical model for using imperfect machine learning models to estimate population parameters from a biased training set. Key results include a deep learning model with likely better performance at predicting suicide risk than trained CVs and the ability to predict whether a texter is 21 or under with 88.4% accuracy. We produce three metrics for conversation success and evaluate the validity and usefulness for each. Finally, reversal of participation bias provides evidence that women, who make up 80.3% of conversations with an associated texter survey, make up closer to 73.5%- 74.8% of all conversations; and that if, after every conversation, the texter had shared whether they found their conversation helpful, affirmative answers would fall from 85.1% to 45.45% - 46.51%.