GEC refreshed: from character diffs to phrase-level review

In August 2026, we updated the GEC frontend with a clearer way to review Chinese corrections. The workspace now uses the Yanlan-expand phrase-expansion model to expand single-character differences into phrase-level spans. Editors can see the wording around a correction directly in the comparison.

From an isolated character to the phrase it belongs to. Yanlan-expand expands the span shown in the diff, helping reviewers judge a suggested change in its linguistic context.

Yanlan-expand: context in every correction

Previously, a diff highlighted individual changed characters. When reviewing a suggestion, the editor had to look back at the sentence to identify the full word or phrase. Yanlan-expand brings that context into the highlighted span: the original wording and its proposed replacement can be read together as a complete unit.

For news and publishing organizations, government communications, and financial or legal editorial teams, this makes it easier to assess how a correction affects the surrounding wording before accepting it.

Illustrative example: “提高效滤” → “提高效率”. The changed character is the same; the highlighted comparison expands from “滤 → 率” to the complete word “效滤 → 效率”. This diagram explains the display behavior, not an evaluation result.

Original text and suggestions, in one workspace

The refreshed layout brings the source text and detection results into a side-by-side review view. Suggestions show their source and status, with controls to locate the original wording, accept a change, or ignore it. Combined with phrase-level diffs, reviewers can assess both the correction and its context.

Sample notice for text review

Interactive text review: select an outlined phrase to inspect, accept, or ignore its correction.

Rich-text review, with the layout in view

The rich-text workspace keeps headings, images, and formatted paragraphs visible while highlighting wording for review. Teams can paste content or upload a .docx document, then check suggestions in the context of the formatted material.

Original article photograph: server room — Interactive rich-text example: formatting and imagery remain visible. Select a red outline or open Suggestions to review two deliberately introduced sample typos, shown as incorrect → corrected.

Video subtitles stay connected to the footage

The subtitle workspace brings video, recognized subtitles, and correction suggestions into the same view. Timestamps link a subtitle issue to its position in the video, so broadcast and media teams can check the wording against the footage.

Sample video frame — Interactive subtitle workspace. Select a timestamp to view its frame, with a red outline around the original error. Inspect, accept, ignore, or reset 9 sample suggestions across 12 subtitle excerpts.

Audio and subtitles, reviewed on separate tracks

The audio/video workspace separates suggestions for the transcribed audio track and the subtitle track. Timestamped segments connect each finding to the media, so reviewers can check spoken content and on-screen subtitles in the same workspace.

Model quality and review clarity

Yanlan V3.0 provides the Chinese correction capability. This August frontend update focuses on how corrections are presented and reviewed, with Yanlan-expand adding phrase context to the diff. Read the V3.0 release for the model evaluation and comparison results.

Read the Yanlan V3.0 release →

The refreshed GEC workspace is available

Publishing, government, and enterprise teams are welcome to contact us for a trial of Yanlan and the updated proofreading workspace.

Contact for trial