<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lee-etal-2026-lost">
    <titleInfo>
        <title>Lost in Morphology at WMT26: A Korean-English Test Suite for LLM Translation Blind Spots</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Deun</namePart>
        <namePart type="given">Sol</namePart>
        <namePart type="family">Lee</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Dae</namePart>
        <namePart type="given">Woong</namePart>
        <namePart type="family">Kim</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yun</namePart>
        <namePart type="given">Ha</namePart>
        <namePart type="family">Seo</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the Eleventh Conference on Machine Translation</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Barry</namePart>
            <namePart type="family">Haddow</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Tom</namePart>
            <namePart type="family">Kocmi</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Philipp</namePart>
            <namePart type="family">Koehn</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Christof</namePart>
            <namePart type="family">Monz</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-474-3</identifier>
    </relatedItem>
    <abstract>We present lost-in-morphology, a 3,536-item Korean to English challenge suite spanning nine phenomena and evaluated on 23 systems submitted to WMT26. The suite targets ellipsis punctuation, stacked particles, wordplay, topic/subject marking, evidentiality, scrambling, auxiliary chains, connective endings and ideophones. We match each phenomenon to its observable evidence: deterministic rules for visible or lexically constrained contrasts, structured LLM observations for semantic features, joint scoring for minimal pairs, distributional metrics for multi-variant frames, and a strategy-aware panel for wordplay. Learned judges and classifiers must pass subset-specific admission tests before their outputs contribute to a system score. Results reveal failures hidden by aggregate MT quality. Ellipsis is usually preserved, but one model family deletes the marked utterance in most of its failures. Particle fidelity falls from 0.95 at stack depth 1 to 0.70 at depth 4, while distinct scalar particles collapse into bare English even. Evidential and auxiliary meanings are often neutralised although the core proposition survives, and wordplay remains the hardest subset: the best system scores 0.385 against 0.546 for the curated references. Judge validation is itself diagnostic: the only conflict-free wordplay judge fails through ceiling saturation, whereas three judges that are also evaluated systems pass without detectable family preference. The suite therefore measures both translation failures and the reliability limits of the instruments used to identify them.</abstract>
    <identifier type="citekey">lee-etal-2026-lost</identifier>
    <location>
        <url>https://aclanthology.org/2026.wmt-main.89/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>1555</start>
            <end>1567</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Lost in Morphology at WMT26: A Korean-English Test Suite for LLM Translation Blind Spots
%A Lee, Deun Sol
%A Kim, Dae Woong
%A Seo, Yun Ha
%Y Haddow, Barry
%Y Kocmi, Tom
%Y Koehn, Philipp
%Y Monz, Christof
%S Proceedings of the Eleventh Conference on Machine Translation
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-474-3
%F lee-etal-2026-lost
%X We present lost-in-morphology, a 3,536-item Korean to English challenge suite spanning nine phenomena and evaluated on 23 systems submitted to WMT26. The suite targets ellipsis punctuation, stacked particles, wordplay, topic/subject marking, evidentiality, scrambling, auxiliary chains, connective endings and ideophones. We match each phenomenon to its observable evidence: deterministic rules for visible or lexically constrained contrasts, structured LLM observations for semantic features, joint scoring for minimal pairs, distributional metrics for multi-variant frames, and a strategy-aware panel for wordplay. Learned judges and classifiers must pass subset-specific admission tests before their outputs contribute to a system score. Results reveal failures hidden by aggregate MT quality. Ellipsis is usually preserved, but one model family deletes the marked utterance in most of its failures. Particle fidelity falls from 0.95 at stack depth 1 to 0.70 at depth 4, while distinct scalar particles collapse into bare English even. Evidential and auxiliary meanings are often neutralised although the core proposition survives, and wordplay remains the hardest subset: the best system scores 0.385 against 0.546 for the curated references. Judge validation is itself diagnostic: the only conflict-free wordplay judge fails through ceiling saturation, whereas three judges that are also evaluated systems pass without detectable family preference. The suite therefore measures both translation failures and the reliability limits of the instruments used to identify them.
%U https://aclanthology.org/2026.wmt-main.89/
%P 1555-1567

Markdown (Informal)

[Lost in Morphology at WMT26: A Korean-English Test Suite for LLM Translation Blind Spots](https://aclanthology.org/2026.wmt-main.89/) (Lee et al., WMT 2026)

• Lost in Morphology at WMT26: A Korean-English Test Suite for LLM Translation Blind Spots (Lee et al., WMT 2026)

ACL

• Deun Sol Lee, Dae Woong Kim, and Yun Ha Seo. 2026. Lost in Morphology at WMT26: A Korean-English Test Suite for LLM Translation Blind Spots. In Proceedings of the Eleventh Conference on Machine Translation, pages 1555–1567, Budapest, Hungary. Association for Computational Linguistics.