Training LLMs to Recognize Hedges in Spontaneous Narratives

TL;DR

This work investigates how to train models to recognize hedges in spontaneous narratives, grounding hedge detection in psycholinguistic theory of collateral signals. It introduces the Roadrunner-Hedge corpus with high-quality, token-level hedge annotations and compares fine-tuned BERT against zero/few-shot prompting of GPT-4o and LLaMA-3, finding that a dedicated fine-tuned model performs best. An error analysis reveals systematic failure modes in LLMs and demonstrates an LLM-in-the-Loop approach to improve annotations and reveal linguistically interesting ambiguities, guiding future research on interpreting and generating collateral signals in conversation. The findings suggest hedging detection is learnable with targeted supervision, but large LLMs alone do not exhibit robust hedge understanding, underscoring the need for specialized training and multi-modal cues to reach genuine conversational agency.

Abstract

Hedges allow speakers to mark utterances as provisional, whether to signal non-prototypicality or "fuzziness", to indicate a lack of commitment to an utterance, to attribute responsibility for a statement to someone else, to invite input from a partner, or to soften critical feedback in the service of face-management needs. Here we focus on hedges in an experimentally parameterized corpus of 63 Roadrunner cartoon narratives spontaneously produced from memory by 21 speakers for co-present addressees, transcribed to text (Galati and Brennan, 2010). We created a gold standard of hedges annotated by human coders (the Roadrunner-Hedge corpus) and compared three LLM-based approaches for hedge detection: fine-tuning BERT, and zero and few-shot prompting with GPT-4o and LLaMA-3. The best-performing approach was a fine-tuned BERT model, followed by few-shot GPT-4o. After an error analysis on the top performing approaches, we used an LLM-in-the-Loop approach to improve the gold standard coding, as well as to highlight cases in which hedges are ambiguous in linguistically interesting ways that will guide future research. This is the first step in our research program to train LLMs to interpret and generate collateral signals appropriately and meaningfully in conversation.