Few Shades of Supervision for Discourse Segmentation

2026-04-19 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

A study by Prevot and Muller evaluates large language model (LLM)-based approaches for segmenting Elementary Discourse Units (EDUs) in spontaneous speech transcripts, a task considered challenging compared to written genres. Utilizing an 8-hour French corpus manually segmented into EDUs, the researchers compared various fine-tuning strategies, including those with weakly supervised labels, against in-context learning methods like few-shot and zero-shot learning. Their findings, published in *Dialogue & Discourse* in December 2025, indicate that classical fine-tuning remains the most effective approach, achieving superior performance with a reasonable amount of gold-annotated data. The research also includes a qualitative analysis, suggesting improvements such as integrating prosodic considerations and handling pauses co-occurring with disfluencies or complex discourse markers.

Key takeaway

For research scientists developing discourse analysis tools for spontaneous speech, you should prioritize classical fine-tuning strategies for large language models. This approach, even with a reasonable amount of gold-annotated data, outperforms in-context learning methods for Elementary Discourse Unit segmentation. Consider incorporating prosodic features and handling disfluencies to further enhance model performance, especially when working with challenging audio transcripts.

Key insights

Classical fine-tuning of LLMs is most effective for EDU segmentation in spontaneous speech.

Principles

• EDUs bridge language grammar and use.
• Spontaneous speech EDU segmentation is hard.
• Gold-annotated data improves fine-tuning.

Method

The study evaluates LLM-based EDU segmentation using an 8-hour French corpus, comparing classical fine-tuning (with and without weak supervision) against few-shot and zero-shot in-context learning.

In practice

• Prioritize classical fine-tuning for EDU tasks.
• Collect reasonable gold-annotated data.
• Consider prosody for speech segmentation.

Topics

• Discourse Segmentation
• Elementary Discourse Units
• Large Language Models
• Fine-tuning
• Spontaneous Speech

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Enhancing Science Classroom Discourse Analysis through Joint Multi-Task Learning for Reasoning-Component Classification — Automated discourse analysis using LLM augmentation improves utterance classification and reveals effective teaching patterns.
• Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models — EBFT fine-tunes language models by matching sequence-level features, improving performance over next-token prediction.
• Think in Sentences: Explicit Sentence Boundaries Enhance Language Model's Capabilities — Explicitly modeling sentence boundaries with delimiters enhances LLM reasoning by inducing structured, sentence-by-sentence processing.
• Building an LLM fine-tuning Dataset — Building multi-speaker Reddit datasets for LLM fine-tuning is complex but achievable, with efficient QLoRA methods requiring minimal samples.
• Specializing a Small Language Model for Closed-Domain Portuguese RAG using Knowledge Graph Supervision — Fine-tuning SLMs with knowledge graph supervision significantly enhances closed-domain RAG performance, surpassing larger LLMs.
• Predicting Causal Effects from Natural Language Queries using Structured Representations — Finetuning and structured representations significantly enhance LLM performance in predicting causal effects from natural language.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.