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

2026-07-02 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

A study on input rewriting with Large Language Models (LLMs) for dialogue discourse parsing (DDP) reveals significant challenges under realistic, unsupervised deployment conditions. Examining three Segmented Discourse Representation Theory (SDRT) datasets and multiple parsers, the research found that last-utterance clarification is considerably less reliable than previously suggested by supervised approaches. Parser-agnostic rewriting frequently introduced more parsing regressions than repairs, as beneficial edits often disrupted critical discourse cues. A "best-of-8" rewriting analysis further indicated that a substantial portion of errors are not fixable through input rewriting alone. While a parser-aware clarifier, trained with GRPO, reduced regressions by up to 37% through conservative abstention, it still failed to consistently improve parsing. These findings reframe clarification as a selective intervention problem, highlighting rewritability prediction as a crucial missing capability for optimizing frozen discourse parsers and broader agentic pipelines.

Key takeaway

For NLP Engineers developing dialogue systems or agentic pipelines, blindly implementing input rewriting for discourse parsing is risky. Your efforts should prioritize developing "rewritability prediction" capabilities to determine when an utterance is genuinely repairable. Without selective intervention, parser-agnostic rewriting can introduce more regressions than repairs, disrupting crucial discourse cues and hindering overall parsing accuracy. Focus on building systems that know when to abstain from rewriting.

Key insights

Input rewriting for dialogue discourse parsing often degrades performance without selective intervention.

Principles

• Parser-agnostic rewriting can harm discourse cues.
• Many DDP errors are not fixable by input rewriting.
• Clarification is a selective intervention problem.

Method

A parser-aware clarifier trained with GRPO learns conservative abstention to reduce regressions by up to 37% in dialogue discourse parsing.

In practice

• Prioritize rewritability prediction for DDP.
• Develop selective intervention strategies for agentic pipelines.

Topics

• Input Rewriting
• Dialogue Discourse Parsing
• Large Language Models
• Agentic Pipelines
• Rewritability Prediction
• SDRT

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• TalkTag: Fine-Grained Morphosyntactic Error Annotation for Transcribed Speech — TalkTag automates fine-grained morphosyntactic error annotation in spoken transcripts using an LLM, even with scarce data.
• BiCon-Gate: Consistency-Gated De-colloquialisation for Dialogue Fact-Checking — BiCon-Gate improves dialogue fact-checking by consistently de-colloquializing claims with contextual semantic gating.
• Readers make targeted regressions to plausible errors in reanalysis of "noisy-channel garden-path" sentences — Readers make targeted eye movements to resolve errors in "noisy-channel garden-path" sentences.
• Linguistic Issues in Computational Linguistics for Non-English Languages — English-centric LLM design creates systematic linguistic weaknesses for non-English languages across multiple levels.
• CR4T: Rewrite-Based Guardrails for Adolescent LLM Safety — Adolescent LLM safety requires transforming unsafe responses into guidance-oriented interactions, not just refusal.
• 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.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.