Utterance Rewriting for Contextual Dialogue Understanding

This direction is merged into another direction: conversational context dependency (contextual reference resolution). Its last version stays readable and citable here.

Version 1 of 117 papers25 conceptsfrozen 2026-10-04

Human conversation is efficient because it is fragmented. We rely on coreference (using "it" or "him" to refer to previous subjects) and ellipsis (omitting words entirely when the meaning is obvious). While these shortcuts make talking easy for humans, they create a "context gap" for AI. If a model only looks at the current turn, a sentence like "I'll take that one too" is semantically empty.

To bridge this gap, researchers have converged on a strategy called Utterance Rewriting. Instead of forcing downstream models to navigate complex multi-turn histories, we use a specialized model to "rewrite" the current turn into a self-contained, standalone sentence that explicitly includes all missing information.

The Core Mechanism: From Fragments to Canonical Sentences

The fundamental goal of an utterance rewriter is to transform a context-dependent turn into a "canonical" form. As Hui Su, Xiaoyu Shen, and their colleagues demonstrated in Improving Multi-turn Dialogue Modelling with Utterance ReWriter, this can be framed as a sequence-to-sequence (seq2seq) task.

The rewriter takes the dialogue history and the current messy utterance as input and produces a "restored" version. To ensure the model doesn't hallucinate, researchers like Chengxi Li and Xiyuan Zhang utilize pointer-generator networks or copy mechanisms. These allow the model to literally "point" to specific entities in the previous turns—like a restaurant name or a specific movie title—and copy them directly into the new sentence. This ensures that the rewritten output remains factually grounded in the conversation.

Why Rewriting Works: Simplifying the Downstream Task

The power of this approach lies in its modularity. By resolving linguistic ambiguities upfront, we turn a difficult multi-turn problem into a much simpler single-turn problem for the rest of the system.

  1. Logical Consistency: Di Jin and Yang Liu showed that chatbots often contradict themselves because they lose track of what "it" refers to. By rewriting utterances to be explicit, they improved contradiction detection by up to 6.5%.
  2. Topic Segmentation: In unsupervised settings, identifying where one topic ends and another begins is hard when sentences are fragmented. Xia Hou and Tongliang Li found that rewriting utterances to resolve pronouns creates a more consistent semantic flow, making topic shifts much easier to detect.
  3. Summarization: Summarizing a chat log is notoriously harder than summarizing a news article. Amanda Bertsch and Graham Neubig introduced a "perspective shift" task that rewrites first-person dialogue into third-person narratives. This transformation allows standard summarization models to work effectively on conversational data.

Scaling and Data

As the field moves toward more complex interactions, the need for high-quality training data has grown. The RiSAWOZ dataset, developed by Jun Quan and Deyi Xiong, provides over 11,000 dialogues with explicit annotations for ellipsis and coreference, specifically to help models learn this rewriting process in task-oriented settings.

Even in long-context scenarios, the intuition of "reconstruction" remains vital. Jia-Nan Li and colleagues' StreamingDialogue uses specialized tokens to compress dialogue, but ensures these tokens are effective by training the model to reconstruct the full, rewritten utterance from the compressed state.

By treating context resolution as a translation task—translating from "human-shorthand" to "machine-explicit"—utterance rewriting provides a robust foundation for more reliable and interpretable conversational AI.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_utterance_rewriting_for_contextual_dialo_243cdf71,
  title = {Utterance Rewriting for Contextual Dialogue Understanding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/utterance-rewriting-for-contextual-dialogue-understanding/rdr_243cdf71dbae4a6b86e86549e430cbe1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_utterance_rewriting_for_contextual_dialo_243cdf71_v1,
  title = {Utterance Rewriting for Contextual Dialogue Understanding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/utterance-rewriting-for-contextual-dialogue-understanding/rdr_243cdf71dbae4a6b86e86549e430cbe1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}