@inproceedings{xu-etal-2026-back,
title = "Back to Basics, Powered by Data: Character-Level Transformers for Robust {C}hinese Machine-Generated Text Detection",
author = "Xu, Shengjie and
Gao, Zhenghao and
Dai, Rujie and
Li, Zijing and
Xie, Xiaoyang and
Lyu, Zehua",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.265/",
pages = "5879--5903",
ISBN = "979-8-89176-439-2",
abstract = "Machine-generated text detection remains difficult in Chinese, where systematic benchmarks and large paired corpora are still scarce. We revisit this problem with a compact 4-layer character-level Transformer, trained on a 363.4-million-character Chinese detection corpus, to our knowledge the largest AI-human paired Chinese resource for this task. The resulting data-centric pipeline attains strong scores on Chinese benchmarks and, for the independently trained English variant, state-of-the-art RAID leaderboard numbers that remain stable under 11 RAID adversarial attack types. Ablations expose a double-edged sword role for punctuation: it eases optimization yet encourages overfitting to surface cues; symbol cleaning shifts the model toward deeper syntactic and discourse patterns and improves generalization. Minimal architecture plus rich supervision thus yields compact detectors suitable as a document-level authorship authenticity screen before downstream Chinese text processing; on a laptop RTX 3070 Ti we measure 5.90 x 10{\textasciicircum}5 tokens/s at under 100 MB VRAM. We present GiantMelon-ZH (Chinese) and GiantMelon-EN (English for RAID): same backbone, disjoint training corpora, transparent data descriptions."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-back">
<titleInfo>
<title>Back to Basics, Powered by Data: Character-Level Transformers for Robust Chinese Machine-Generated Text Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shengjie</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenghao</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rujie</namePart>
<namePart type="family">Dai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zijing</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyang</namePart>
<namePart type="family">Xie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zehua</namePart>
<namePart type="family">Lyu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Machine-generated text detection remains difficult in Chinese, where systematic benchmarks and large paired corpora are still scarce. We revisit this problem with a compact 4-layer character-level Transformer, trained on a 363.4-million-character Chinese detection corpus, to our knowledge the largest AI-human paired Chinese resource for this task. The resulting data-centric pipeline attains strong scores on Chinese benchmarks and, for the independently trained English variant, state-of-the-art RAID leaderboard numbers that remain stable under 11 RAID adversarial attack types. Ablations expose a double-edged sword role for punctuation: it eases optimization yet encourages overfitting to surface cues; symbol cleaning shifts the model toward deeper syntactic and discourse patterns and improves generalization. Minimal architecture plus rich supervision thus yields compact detectors suitable as a document-level authorship authenticity screen before downstream Chinese text processing; on a laptop RTX 3070 Ti we measure 5.90 x 10⌃5 tokens/s at under 100 MB VRAM. We present GiantMelon-ZH (Chinese) and GiantMelon-EN (English for RAID): same backbone, disjoint training corpora, transparent data descriptions.</abstract>
<identifier type="citekey">xu-etal-2026-back</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.265/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>5879</start>
<end>5903</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Back to Basics, Powered by Data: Character-Level Transformers for Robust Chinese Machine-Generated Text Detection
%A Xu, Shengjie
%A Gao, Zhenghao
%A Dai, Rujie
%A Li, Zijing
%A Xie, Xiaoyang
%A Lyu, Zehua
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F xu-etal-2026-back
%X Machine-generated text detection remains difficult in Chinese, where systematic benchmarks and large paired corpora are still scarce. We revisit this problem with a compact 4-layer character-level Transformer, trained on a 363.4-million-character Chinese detection corpus, to our knowledge the largest AI-human paired Chinese resource for this task. The resulting data-centric pipeline attains strong scores on Chinese benchmarks and, for the independently trained English variant, state-of-the-art RAID leaderboard numbers that remain stable under 11 RAID adversarial attack types. Ablations expose a double-edged sword role for punctuation: it eases optimization yet encourages overfitting to surface cues; symbol cleaning shifts the model toward deeper syntactic and discourse patterns and improves generalization. Minimal architecture plus rich supervision thus yields compact detectors suitable as a document-level authorship authenticity screen before downstream Chinese text processing; on a laptop RTX 3070 Ti we measure 5.90 x 10⌃5 tokens/s at under 100 MB VRAM. We present GiantMelon-ZH (Chinese) and GiantMelon-EN (English for RAID): same backbone, disjoint training corpora, transparent data descriptions.
%U https://aclanthology.org/2026.aacl-long.265/
%P 5879-5903
Markdown (Informal)
[Back to Basics, Powered by Data: Character-Level Transformers for Robust Chinese Machine-Generated Text Detection](https://aclanthology.org/2026.aacl-long.265/) (Xu et al., AACL-IJCNLP 2026)
ACL