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

🕒 Published (v1): 2026-07-02 09:57 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper stress-tests zero-shot and RL-based input rewriting as a way to improve frozen dialogue discourse parsers (DDP) without supervised clarification data. Parser-agnostic rewriting consistently introduces more regressions than repairs, and a GRPO-trained parser-aware policy reduces regressions by up to 37% by learning to abstain—but still fails to achieve per-instance selectivity. The key missing capability is rewritability prediction: knowing before intervention whether a given utterance is even repairable.

Problem¶

Prior work (Fan et al., 2025) showed that a supervised LLM clarifier can rewrite ambiguous last utterances to improve incremental SDRT discourse parsing, but such supervision is costly and parser-specific. This paper asks whether last-utterance clarification remains effective under realistic deployment conditions where no clarification supervision exists and the clarifier must rely on zero-shot prompting or frozen-parser feedback alone.

Method¶

Parser-agnostic clarification. Seven zero-shot rewriting strategies of increasing intervention depth are applied to the final utterance before passing the modified context to a frozen parser: L0-TYPO (typo correction), L1-NORM (lexical normalization), L2-EXPL (ellipsis completion), L3-COREF (coreference grounding), L4-CONN (discourse connective insertion), L5-MIX (any subset of L0–L4), and L6-FREE (unconstrained free-form rewriting). Strategy-specific prompts are selected by maximizing BERTScore meaning-preservation on a validation set.

Parser-aware clarification. A rewriting policy \(\pi_\phi\) (Qwen2.5-7B-Instruct) is trained with GRPO using a verifiable reward derived entirely from the frozen parser's behavior. At each dialogue step the policy generates the original or a rewritten utterance \(\tilde{u}_t\); the reward combines a format component (\(r_{fmt} \in \{-2, +1\}\)) and a parsing component (\(+2\) for fixing an error, \(-2\) for introducing one, \(+1\) for partial fix, \(0\) otherwise). No clarification supervision is used.

Evaluation. Micro-F1 for link attachment and full structure (Link+Rel) on three SDRT datasets (STAC, Molweni, MSDC) against Qwen3-8B/14B LoRA-finetuned generative parsers and the SDDP discriminative graph parser.

Repairability ceiling. A best-of-8 oracle baseline samples 8 independent L6-FREE rewrites per erroneous prediction and marks examples repairable if any rewrite corrects the error.

Key Contributions¶

Systematic evaluation of seven unsupervised clarification strategies across three SDRT datasets and four parsers, establishing that parser-agnostic rewriting is net-negative.
Demonstration that GRPO-trained parser-aware RL reduces regressions by up to 37% (STAC) but converges to dataset-level intervention regimes (abstain vs. rewrite-everything) rather than per-instance selective policies.
Repairability ceiling analysis showing ~72–80% of discourse parsing errors are structurally not fixable through any input rewrite, reframing clarification as a selective intervention problem.
Identification of rewritability prediction — pre-intervention classification of whether an utterance error is repairable — as the key missing capability for input-side optimization of frozen NLP components.

Results¶

Parser-agnostic (Qwen3-8B, STAC): L4-CONN causes 21 repairs vs. 88 regressions; L6-FREE causes 17 vs. 38; Link+Rel F1 drops from 0.598 baseline in most conditions. The conservative L0-TYPO is near-neutral (2 repairs, 3 regressions) but provides no meaningful gain.
Parser-agnostic (SDDP): Same pattern holds for the discriminative parser; L4-CONN produces the largest degradation across all three datasets.
Parser-aware RL vs. L6-FREE (STAC): regressions 38 → 24 (−37%), repairs 17 → 7; Edit% drops from 49.9% to 21.4%; Link+Rel F1 improves (0.577 → 0.586). Gains come from abstention, not better rewrites.
Parser-aware RL (Molweni): Edit% reaches 98.8%; regressions fall only 128 → 113; Link+Rel F1 drops vs. L6-FREE baseline. Policy collapses to near-universal rewriting.
Parser-aware RL (MSDC): regressions increase (105 → 125) vs. L6-FREE; policy is still net-harmful despite declining Edit%.
Repairability ceiling: 80.0% / 72.5% / 71.8% of errors on STAC / Molweni / MSDC are unrepaired by all 8 sampled rewrites; "highly repairable" cases represent only 2.4% / 7.2% / 4.6% of errors.
RL repair rate by repairability bucket: near 0% for "not repairable" examples; up to 31.9% (Molweni) and 23.5% (MSDC) for "highly repairable" bucket.

Limitations¶

Parser-aware clarifier uses Qwen2.5-7B-Instruct (not Qwen3) due to RL framework compatibility constraints, introducing a clarifier–parser model mismatch that could affect results.
RL training is conducted against only the Qwen3-8B parser; generalization to other parsers is not directly evaluated for the RL condition.
The best-of-8 repairability estimate is an approximation; with a stronger or more diverse rewriter, repairability fractions could differ.
Scope is restricted to SDRT-based DDP; generalization to RST, PDTB, or other discourse formalisms is not tested.
The paper does not propose or evaluate any concrete rewritability prediction model, leaving the identified key direction as future work.
Incremental setting means errors compound and past decisions cannot be revised, which may make clarification harder than in batch settings.

Relevance to Agentic AI / LLM Agents¶

This paper directly probes a fundamental assumption in agentic pipeline design: that an upstream LLM module (rewriter, planner, tool orchestrator) can improve a frozen downstream component through prompt-driven input transformation. The finding that generic, non-selective rewriting is net-negative—and that RL-trained agents learn coarse intervention regimes rather than fine-grained selectivity—is a concrete empirical caution against deploying unrestricted LLM preprocessors in front of specialized frozen tools. The concept of rewritability prediction maps directly onto the broader agentic problem of deciding when to invoke a tool or subagent versus abstaining, which is a critical yet understudied capability in current LLM agent scaffolding. The GRPO-based parser-aware training setup is also a clean instance of the feedback-driven agent adaptation pattern common in agentic RL research.