The study reported a large gap in misspelling rates during a live traffic test of a system for correcting short Chinese search queries, alongside a small increase in overall search volume. In the bucketed comparison, the reported rate was 2.58% for the base arm and 0.86% for the experiment arm, alongside a 0.122% lift in search volume. The figures come from a preprint, and the study does not report exact arm sizes or uncertainty estimates.
The system, called GUIDE, is designed for situations in which short queries are difficult to interpret and new words appear quickly. The paper evaluates it as an unsupervised approach, using reconstruction and search-frequency information instead of relying on manually corrected examples for every query.
A correction system built around sound and shape
GUIDE groups Chinese characters that can be confused because they sound alike or look alike, giving characters in each group a shared identifier. An encoder-decoder Transformer then uses those representations to reconstruct the original character sequence.
Its training objective also weights queries by how often they are searched and how recently they appeared. That design shifts learning toward frequent, fresh queries while retaining the shared-identifier reconstruction task, an approach aimed at changing search vocabularies.
The offline tests used query-level precision, recall and F1. In plain terms, those measures examine the accuracy of proposed corrections, the errors the system catches and a combined score. QSpell 250K lists 200,000 training queries and 50,000 test queries. KwaiSearch lists 180 million training items and 30,000 test items, although its training error count is marked as unknown.
Results varied across the two benchmarks
On QSpell 250K, the reported six-layer GUIDE result had a precision of 0.4367, a recall of 0.5547 and an F1 score of 0.4887. On KwaiSearch, the same reported row reached 0.6821 precision, 0.8419 recall and 0.7536 F1. The paper reports these scores without confidence intervals, p-values or another formal significance test.
The comparison of character-grouping strategies points to a benefit from combining the two signals. The combined phonetic-and-visual setting had the highest listed F1 among the variants, at 0.4730 on QSpell 250K and 0.7508 on KwaiSearch.
A separate KwaiSearch ablation, using a three-layer version, showed progressively higher F1 across the reported training objectives: 0.5947 with uniform weighting, 0.7048 with frequency-only weighting and 0.7508 with the full time-decay and frequency objective.
The live test has a reporting wrinkle
The online study ran a bucketed A/B test, comparing a base arm with an experiment arm, from Sept. 13 to Sept. 22, 2025. It covered 4.2% of production main-search traffic for 10 days. Misspelling rates were estimated by randomly sampling online queries and manually annotating them.
Elsewhere, the online-results section reports a different comparison: an 80.1% relative decrease in misspelling rate, from 2.01% to 0.40%. It also lists increases in page views from several search entry points, ranging from 0.152% to 0.986%, and again gives a 0.122% overall search-volume increase. The preprint does not explain how this second set of figures relates to the 2.58% and 0.86% rates reported for the bucketed test.
Useful evidence, but a narrow test
The method's stated scope is length-preserving character substitution. It does not directly address insertion, deletion, word segmentation or phrase-level rewriting, so the reported findings do not establish performance on those other types of query error.
The online comparison also leaves important details unspecified. Exact arm sizes, allocation procedures and formal uncertainty estimates are not reported, making the size and reliability of the operational differences harder to assess.
Kuaishou Technology is acknowledged as supporting the work. The document is an arXiv preprint identified as arXiv:2608.25343v1 and dated Aug. 26, 2026. The paper states that KwaiSearch is available through a GitHub address containing the word PLACEHOLDER, but the supplied text does not provide a verifiable working link.
Paper data and sources
Original title: GUIDE: Generative Unsupervised Chinese Query Correction via Phonetic and Visual Shared-ID Encoding
Authors: Lei Yang, Binbin Huang, Jiwei Tan et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text