How did we get here? Summarizing conversation dynamics

This is the official repo for our paper How did we get here? Summarizing conversation dynamics.

Yilun Hua, Nicholas Chernogor, Yuzhe Gu, Seoyeon Julie Jeong, Miranda Luo, Cristian Danescu-Niculescu-Mizil

Data

First Release (the data used in our paper):

• 437 conversations and their Summaries of Conversation Dynamics (SCDs) generated by our GPT procedural prompt.
• 50 of the conversations in the test set also have corresponding human-written SCDs.
• The same data can also be accessed via the ConvoKit library, which contains additional conversation meta-data and various useful functionalities.

We are expanding our dataset. Future releases will include more human-written SCDs and conversation-level dynamics labels to facilitate future research and broader applications. Please stay tuned.

Experiment pipelines

Please refer to src/summary_generation_derailment_prediction for our core experiments. Please refer to src/informativeness_check_questions for our informativeness check.