<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wijesiri-etal-2026-confident">
    <titleInfo>
        <title>Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Isuru</namePart>
        <namePart type="family">Wijesiri</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Nisansa</namePart>
        <namePart type="family">de Silva</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Kavindu</namePart>
        <namePart type="family">Warnakulasuriya</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Aloka</namePart>
        <namePart type="family">Fernando</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Surangika</namePart>
        <namePart type="family">Ranathunga</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>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</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">979-8-89176-462-0</identifier>
    </relatedItem>
    <abstract>Automatic Post-Editing (APE) for low-resource languages (LRLs) often fails to improve Machine Translation (MT), and the score alone cannot say why: whether more training would help, or whether the training data is too inconsistent to learn from. We introduce a black-box, inference-time diagnostic that tells these two cases apart without retraining or annotation. It varies an edit-distance penalty łambda that drives the model from free editing towards copying the MT, and reads two signals: (1) the shape of the Translation Edit Rate (TER)-vs-łambda curve, U-shaped if edits from the model reduce error and monotonically decreasing if none does; and (2) the ordering of constraint variants that trust model confidence to increasing degrees, which shows whether confidence tracks edit quality. Across decoder-only and encoder–decoder models on English–Sinhala, the diagnostic exposes two failure modes consistent with a heterogeneous post-edit signal as the underlying cause: Binary Collapse, where the model copies the MT or makes off-target edits, and Confident Miscalibration, where the confidence signals we test do not separate useful edits from unnecessary ones. The pattern holds on English–Marathi and English–Tamil, with the failure modes tracking the post-edit distribution rather than MT quality or language family. Beyond diagnosis, the curve shape prescribes a concrete next step for practitioners; in the favorable case, a static constraint yields a free inference-time accuracy gain. We release the first English–Sinhala (\sim66k) and a new English–Tamil (\sim39k) APE datasets with all code.</abstract>
    <identifier type="citekey">wijesiri-etal-2026-confident</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1079/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>22838</start>
            <end>22866</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing
%A Wijesiri, Isuru
%A de Silva, Nisansa
%A Warnakulasuriya, Kavindu
%A Fernando, Aloka
%A Ranathunga, Surangika
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F wijesiri-etal-2026-confident
%X Automatic Post-Editing (APE) for low-resource languages (LRLs) often fails to improve Machine Translation (MT), and the score alone cannot say why: whether more training would help, or whether the training data is too inconsistent to learn from. We introduce a black-box, inference-time diagnostic that tells these two cases apart without retraining or annotation. It varies an edit-distance penalty łambda that drives the model from free editing towards copying the MT, and reads two signals: (1) the shape of the Translation Edit Rate (TER)-vs-łambda curve, U-shaped if edits from the model reduce error and monotonically decreasing if none does; and (2) the ordering of constraint variants that trust model confidence to increasing degrees, which shows whether confidence tracks edit quality. Across decoder-only and encoder–decoder models on English–Sinhala, the diagnostic exposes two failure modes consistent with a heterogeneous post-edit signal as the underlying cause: Binary Collapse, where the model copies the MT or makes off-target edits, and Confident Miscalibration, where the confidence signals we test do not separate useful edits from unnecessary ones. The pattern holds on English–Marathi and English–Tamil, with the failure modes tracking the post-edit distribution rather than MT quality or language family. Beyond diagnosis, the curve shape prescribes a concrete next step for practitioners; in the favorable case, a static constraint yields a free inference-time accuracy gain. We release the first English–Sinhala (\sim66k) and a new English–Tamil (\sim39k) APE datasets with all code.
%U https://aclanthology.org/2026.findings-emnlp.1079/
%P 22838-22866

Markdown (Informal)

[Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing](https://aclanthology.org/2026.findings-emnlp.1079/) (Wijesiri et al., Findings 2026)

• Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing (Wijesiri et al., Findings 2026)

ACL

• Isuru Wijesiri, Nisansa de Silva, Kavindu Warnakulasuriya, Aloka Fernando, and Surangika Ranathunga. 2026. Confident but Wrong: A Constrained Decoding Diagnostic for Low-Resource Automatic Post-Editing. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 22838–22866, Budapest, Hungary. Association for Computational Linguistics.