Beyond Supervised Clarification: Input Rewriting with LLMs for Dialogue Discourse Parsing

2026-07-20 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A study presented at the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue in August 2026 investigates the effectiveness of input rewriting with Large Language Models (LLMs) for Dialogue Discourse Parsing (DDP) under realistic, unsupervised conditions. Unlike prior supervised clarification models that improved DDP accuracy by rewriting fragmentary utterances, this research finds that last-utterance clarification is significantly less reliable without supervision. Across three Segmented Discourse Representation Theory (SDRT) datasets and multiple parsers, parser-agnostic rewriting frequently introduces more regressions than repairs, as edits disrupt crucial discourse cues. A best-of-8 rewriting analysis revealed that a large fraction of errors are not repairable through input rewriting alone. While a parser-aware clarifier trained with GRPO reduced regressions by up to 37%, it still failed to consistently improve parsing. The authors identify "rewritability prediction"—determining if an utterance is repairable before intervention—as the critical missing capability for optimizing frozen discourse parsers and enhancing agentic pipelines.

Key takeaway

For NLP Engineers implementing input rewriting with LLMs for dialogue discourse parsing or agentic pipelines, you should recognize that unsupervised clarification often introduces more regressions than repairs. Blindly applying input rewriting without a "rewritability prediction" mechanism risks degrading your frozen parser's performance. Prioritize developing or integrating systems that can selectively intervene, determining if an utterance is genuinely repairable before attempting modification, to ensure consistent parsing improvements.

Key insights

Unsupervised input rewriting for dialogue discourse parsing often causes regressions, highlighting the need for rewritability prediction.

Principles

• Parser-agnostic rewriting can disrupt discourse cues.
• Many DDP errors are not repairable via input rewriting.
• Selective intervention is crucial for input-side optimization.

Method

A parser-aware clarifier was trained using GRPO to learn conservative abstention, aiming to reduce negative impacts.

Topics

• Dialogue Discourse Parsing
• LLM Input Rewriting
• Zero-shot Prompting
• SDRT Datasets
• Agentic Pipelines
• Rewritability Prediction

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

Related on AIssential

• Beyond Supervised Clarification: Input Rewriting with LLMs for Dialogue Discourse Parsing — Input rewriting for dialogue discourse parsing often degrades performance without selective intervention.
• When the Database Fails: Prompting LLM Dialogue Agents for Safe Recovery in Task-Oriented Dialogue — Prompting with structured database status significantly reduces LLM hallucination in task-oriented dialogue without retraining.
• Most LLM Conformity Needs No Speaker: Measuring the Speaker-Free Floor in Peer-Pressure Benchmarks — LLM "conformity" is largely driven by repeated assertions, not explicit social influence, requiring a speaker-free baseline.
• From Dialogue to Learner Modeling: Identifying Candidate Signals of Productive Use in LLM-Based Grammar Practice — Learner production in LLM-based dialogue tutors provides valuable, fine-grained signals for adaptive grammar practice and progress modeling.
• Short-form Text Rewriting with Phi Silica — Targeted adaptation of Small Language Models (SLMs) can significantly improve their performance for precision-critical short-form text rewriting tasks.
• Diagnosis Is Not Prescription: Linguistic Co-Adaptation Explains Patching Hazards in LLM Pipelines — Correcting LLM agent bottlenecks directly can degrade performance due to downstream modules' implicit adaptation to upstream errors.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

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