Human vs LLM in Conversational Repair Annotation: A New Resource and Comparative Study

Abstract

Addressing the scarcity of annotated data for Other-Initiated Repair (OIR), when recipients interrupt conversation progressivity to signal trouble, prompting speakers to provide repair, this work introduces OIR annotations for the NOXI corpus, achieving considerable reliability. We evaluate whether LLMs can reliably annotate OIR sequences using structured Chain-of-Thought prompting and conduct comparative analysis across two corpora: NOXI (natural dialogue) and CABB-S (Dutch, task-oriented), finding weak alignment between LLMs and human annotations, particularly in recognizing trouble-signaling. Analyzing human-LLM disagreement using the LLM-generated explanations revealed limitations: models rely on lexical patterns rather than conversational context, construct reasonable-sounding but misleading narratives, highlighting crucial limitations for both automated annotation of complex interactional phenomena.

Resources

Details