Fine-Grained Taxonomies for Document Revision Intent

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 259 papers78 conceptsfrozen 2026-10-04

When a human editor revises a text, they rarely think in terms of character offsets or word deletions. Instead, they think in terms of intent: "this sentence is too wordy," "this fact is outdated," or "this tone is too aggressive." However, for decades, Natural Language Processing (NLP) has relied on mechanical metrics like Levenshtein distance or BLEU scores to measure change. These metrics are "blunt instruments," as Rémi Cardon and colleagues argue in Linguistic Corpus Annotation for Automatic Text Simplification Evaluation, because they mask the specific linguistic strengths and weaknesses of a system.

The research direction emerging across these papers is a shift toward fine-grained revision taxonomies. By moving from what changed to why it changed, researchers are building models that can finally understand the iterative nature of human writing.

From Mechanical Edits to Logical Actions

The fundamental problem with traditional edit distance is that it treats all changes as equal. A typo fix and a complete rhetorical reframing might both involve changing five characters, but their "logical" weight is vastly different.

Holger Schwenk and Frédéric Blain addressed this by introducing the Post-Editing Action (PEA) in Qualitative Analysis of Post-Editing for High Quality Machine Translation. Instead of counting keystrokes, they categorized edits into linguistic units like "Noun Phrase fixes." This shift revealed that a tiny fraction of terminological errors often accounts for a massive portion of human editing effort.

However, identifying the action is only half the battle. As Somayajulu Sripada and Ehud Reiter noted in their evaluation of NLG systems, automated tools can detect that a change occurred but cannot determine the underlying motivation. Without knowing if a change was a "critical factual correction" or a "subjective stylistic improvement," developers cannot accurately prioritize system updates.

The Multi-Layered Taxonomy of Intent

To bridge this gap, researchers have developed hierarchical frameworks that categorize edits across multiple dimensions. A recurring pattern in the work of Iryna Gurevych and Johannes Daxenberger is the use of tiered classification systems. In Automatically Classifying Edit Categories in Wikipedia Revisions, they proposed a 21-category taxonomy that distinguishes between:

  1. Surface-level edits: Fluency, grammar, and spelling.
  2. Semantic-level edits: Adding facts, modifying claims, or changing evidence.
  3. Structural/Administrative edits: Reordering paragraphs or formatting.

This distinction is crucial because different types of writing prioritize different edits. In scientific writing, Chao Jiang and Wei Xu found that authors use ARXIVEDITS to manage "statement strength" and clarity. Meanwhile, in journalism, Alexander Spangher showed in NewsEdits that edits often track the "shelf life" of a fact, evolving as new information surfaces.

Decoupling Action from Communication

The most recent advancement in this field is the realization that the mechanical action (e.g., "deletion") and the communicative intent (e.g., "removing a weak claim") should be modeled separately.

Qian Ruan and Ilia Kuznetsov formalized this in the Re3 framework. Their research suggests that while substantive additions and deletions are typically used to manage facts, modifications to existing text are primarily focused on improving grammar. By decoupling these, models can better summarize why a document changed—moving from a "diff" view to a "thematic summary" view, as explored by T. Y. S. S. Santosh in A Tale of Two Revisions.

Why This Works: Semantic Priors

The reason these taxonomies improve AI performance is that they act as semantic priors. When a model like Dheeraj Rajagopal’s system in One Document, Many Revisions is told the intent of an edit before generating a description, it is significantly less likely to "hallucinate" irrelevant details. The intent provides a conceptual anchor that constrains the model's output to the actual logic of the transformation.

This has profound implications for the future of AI writing assistants. Instead of just suggesting a "better" sentence, future systems will be able to follow complex, knowledge-intensive instructions—like those in the KIWI dataset by Fangyuan Xu—to iteratively refine a document's narrative, not just its syntax. By understanding the "why" behind the "what," we move from simple text correction to true collaborative authorship.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 42 papers
  2. Version 1 · 2026-10-04 · 59 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_revision_intent_classification_fd49d55c,
  title = {revision intent classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/revision-intent-classification/rdr_fd49d55c55424e949973a53727edb323},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_fine_grained_taxonomies_for_document_rev_fd49d55c_v1,
  title = {Fine-Grained Taxonomies for Document Revision Intent},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fine-grained-taxonomies-for-document-revision-intent/rdr_fd49d55c55424e949973a53727edb323/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}