Mahesh Babu Kommalapati, Xiao Gu, Harshit Pandey, Christie J. Rizzo, Charlene Collibee, Silvio Amir, Aarti Sathyanarayana | February 2025

Abstract: Intimate Partner Violence (IPV) among ado- lescents is a major public health concern, particularly for justice-involved adolescents which are at higher risk of experiencing IPV. Early detection of disagreements between romantic partners can provide an opportunity for just-in-time interventions to prevent escalation. Prior work has proposed methods for early detection of disagreements based on metadata features of text message conversations. In this work, we build on these prior efforts and investigate the impact of explicitly modeling the contents of text conversations for disagreement detection. We develop and evaluate supervised classifiers that combine metadata features with sentiment and semantic features of texts and compare their performance against few- shot learning with instruction-tuned Large Language Models (LLMs). We conduct experiments on a dataset collected to study the communication patterns and risk factors associated with IPV among justice-involved adolescents. In addition, we measure models’ generalization to out-of-distribution samples using an external dataset comprising adolescents enrolled in child welfare services. We find that: (i) text-based features improve predictive performance but do not help models generalize to other populations; and (ii) LLMs struggle in this setting but can outperform supervised classifiers in out- of-distribution samples.