@inproceedings{xiao-chao-2026-lexically,
title = "Lexically Stable but Semantically Improvable: On the Editability of Machine Translation Drafts for {LLM} Post-Editing",
author = "Xiao, Tianlin and
Chao, WenHan",
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.58/",
pages = "1236--1259",
ISBN = "979-8-89176-439-2",
abstract = "Large language models (LLMs) perform strongly in machine translation, but their effectiveness as translation refiners depends on the draft source. We compare direct translation, self-refinement, and cross-system refinement on English-to-Chinese WMT test sets. Sockeye drafts obtain higher lexical-overlap scores but lower COMET than the stronger Qwen3 drafts, and receive larger gains after Qwen3 refinement. Additional controls qualify this pattern. Qwen3-0.6B drafts obtain the largest raw gain but lower final quality, while Qwen3-1.7B drafts start near Sockeye in COMET yet benefit less. Cross-family GLM-4-9B-Chat drafts also improve with Qwen3-8B, but achieve smaller gains and lower final quality than refined Sockeye drafts. Descriptive bucket, common-support, and regression analyses suggest that initial quality alone is an incomplete account of these differences. xCOMET-XL score and span diagnostics further characterize the contrast with Qwen3-8B self-refinement. We frame translation editability as a property of the evaluated draft{--}editor pairing, assessed jointly through quality gain, improvement probability, degradation risk, and final quality."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xiao-chao-2026-lexically">
<titleInfo>
<title>Lexically Stable but Semantically Improvable: On the Editability of Machine Translation Drafts for LLM Post-Editing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tianlin</namePart>
<namePart type="family">Xiao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">WenHan</namePart>
<namePart type="family">Chao</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>Large language models (LLMs) perform strongly in machine translation, but their effectiveness as translation refiners depends on the draft source. We compare direct translation, self-refinement, and cross-system refinement on English-to-Chinese WMT test sets. Sockeye drafts obtain higher lexical-overlap scores but lower COMET than the stronger Qwen3 drafts, and receive larger gains after Qwen3 refinement. Additional controls qualify this pattern. Qwen3-0.6B drafts obtain the largest raw gain but lower final quality, while Qwen3-1.7B drafts start near Sockeye in COMET yet benefit less. Cross-family GLM-4-9B-Chat drafts also improve with Qwen3-8B, but achieve smaller gains and lower final quality than refined Sockeye drafts. Descriptive bucket, common-support, and regression analyses suggest that initial quality alone is an incomplete account of these differences. xCOMET-XL score and span diagnostics further characterize the contrast with Qwen3-8B self-refinement. We frame translation editability as a property of the evaluated draft–editor pairing, assessed jointly through quality gain, improvement probability, degradation risk, and final quality.</abstract>
<identifier type="citekey">xiao-chao-2026-lexically</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.58/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>1236</start>
<end>1259</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Lexically Stable but Semantically Improvable: On the Editability of Machine Translation Drafts for LLM Post-Editing
%A Xiao, Tianlin
%A Chao, WenHan
%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 xiao-chao-2026-lexically
%X Large language models (LLMs) perform strongly in machine translation, but their effectiveness as translation refiners depends on the draft source. We compare direct translation, self-refinement, and cross-system refinement on English-to-Chinese WMT test sets. Sockeye drafts obtain higher lexical-overlap scores but lower COMET than the stronger Qwen3 drafts, and receive larger gains after Qwen3 refinement. Additional controls qualify this pattern. Qwen3-0.6B drafts obtain the largest raw gain but lower final quality, while Qwen3-1.7B drafts start near Sockeye in COMET yet benefit less. Cross-family GLM-4-9B-Chat drafts also improve with Qwen3-8B, but achieve smaller gains and lower final quality than refined Sockeye drafts. Descriptive bucket, common-support, and regression analyses suggest that initial quality alone is an incomplete account of these differences. xCOMET-XL score and span diagnostics further characterize the contrast with Qwen3-8B self-refinement. We frame translation editability as a property of the evaluated draft–editor pairing, assessed jointly through quality gain, improvement probability, degradation risk, and final quality.
%U https://aclanthology.org/2026.aacl-long.58/
%P 1236-1259
Markdown (Informal)
[Lexically Stable but Semantically Improvable: On the Editability of Machine Translation Drafts for LLM Post-Editing](https://aclanthology.org/2026.aacl-long.58/) (Xiao & Chao, AACL-IJCNLP 2026)
ACL