Ground-truth speech acts for learning indirect intents

Determine whether manually extracted, ground-truth speech acts provide sufficiently clean pragmatic supervision for the conversational derailment forecasting models to learn nuanced indirect and hidden speaker intents.

Background

The proposed approach uses speech acts extracted by the gpt-oss-120b LLM as an auxiliary training signal. Although the extraction shows moderate agreement with human labels for fine-grained classes and almost perfect agreement for broad categories, the paper notes that LLMs may fail to reliably identify indirect intents.

The authors explicitly leave unresolved whether replacing noisy language-model-derived labels with manually extracted ground-truth speech acts would enable the downstream forecasting model to learn nuanced, indirect, or hidden intents rather than merely using speech acts as a compact pragmatic representation.