Michelle Gelman

Home

❯

Conversational Derailment in Dispute Resolution

Conversational Derailment in Dispute Resolution

Nov 04, 20252 min read

  • PyTorch
  • NLP
  • MLFlow
  • RayTune
  • explainability

Conversational Derailment in Dispute Resolution

  • This repository contains a tailored implementation of the Conversational Recurrent Architecture for ForecasTing (CRAFT) neural model, originally introduced in the EMNLP 2019 paper Trouble on the Horizon: Forecasting the Derailment of Online Conversations as they Develop to the dispute resolution domain. We pre-train the CRAFT model architecture with a custom corpus of CaSiNo, Deal no Deal, and KODIS dialogs and fine-tune on the KODIS dataset to research whether we can learn unsupervised representation of conversational dynamics in negotiation-based dialogues and exploit the structure via supervised learning for predicting for outcomes in Dispute resolution (KODIS).

Github Project

PyTorch NLP MLFlow RayTune

CRAFT NLP Model: Evaluating Utterance Exposure and Imbalance Strategies on fine-tuning performance for derailment prediction

  • This project analyzes the model sensitivity to utterance variations and imbalance handling on 9 model variants. We determine whether CRAFT is a good predictor of conversation-level derailment on KODIS using Conversations Gone Awry Dataset (CGA-WIKI)as a baseline comparison. We compares variants using standard classification metrics (F1, AUC, Calibration Curves) at each model’s Youden-optimized threshold on the same Ground test set (CGA-WIKI), with additional diagnostics (horizons, distributions, frustration correlation, token analysis) to interpret differences.
  • Including the submit agreement utterance makes KODIS fine-tuned models more prone to forecasting similar derailment scores for all conversations leading to missed conversation dynamics

Presentation | Github

PyTorch NLP explainability


Graph View

  • Github Project
  • Presentation | Github

Backlinks

  • Data Science
  • HCI
  • About Me

Created with Quartz v4.5.2 © 2025

  • GitHub
  • Discord Community