Goal

To find an unsupervised method to train a model for evaluating the coherence of motivational interviewing dialogues in the case of limited data.

Contributions

• Modified the original dataset and rebuilt a dataset for pairwise ranking.
• Developed and trained a GPT-based model utilizing the pairwise ranking approach, enhancing the performance of generators in motivational interviewing.
• Currently engaged in fine-tuning the generator employing techniques akin to interactive generation, leveraging coherence scores as a surrogate for human feedback.