<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-lin-2026-fuzhoubench">
    <titleInfo>
        <title>FuzhouBench: A Benchmark for Eastern Min Chinese with Tone Sandhi Annotations</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Xiaocan</namePart>
        <namePart type="family">Li</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Haoyu</namePart>
        <namePart type="family">Lin</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 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>Eastern Min Chinese (Fuzhou dialect, ISO 639-3 ‘cdo‘) has roughly ten million speakers yet no dedicated paper at any major NLP venue. Four properties make Fuzhou structurally distinct from Mandarin and challenging for current NLP: (1) a tone-sandhi system conditioned on the following syllable’s citation tone, (2) initial-consonant assimilation in non-final positions, (3) systematic literary-colloquial diglossia (45% of frequent characters carry \geq 2 distinct segmental readings), and (4) substantial lexical divergence (everyday concepts use entirely different morphemes from Mandarin). We release **FuzhouBench**, the first public NLP benchmark for Eastern Min, built on a community-maintained open dictionary: 22,890 word entries with 51,505 standardized Yngping romanizations, 11,276 explicit tone-sandhi annotations, 5,160 explanations, and 3.06 hours of paired dictionary-style audio from two native speakers. We define three tasks (grapheme-to-phoneme, sandhi prediction, definition retrieval) with full baselines. Frontier LLMs score under 3% word-exact on few-shot Yngping G2P, while a LoRA fine-tune on the released pairs reaches 17.2%: learnable, but far from saturated. A symbolic sandhi rule learned from the training pool beats every LLM by up to 18 points; and BM25 collapses from MRR 0.799 to 0.372 when queries shift from dictionary-extracted to colloquial Mandarin, partially recoverable via a two-stage hybrid + cross-encoder pipeline (MRR 0.716). Data, evaluation splits, and code are available at https://github.com/Nang-muoi-sing/FuzhouBench.</abstract>
    <identifier type="citekey">li-lin-2026-fuzhoubench</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.618/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>14194</start>
            <end>14208</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T FuzhouBench: A Benchmark for Eastern Min Chinese with Tone Sandhi Annotations
%A Li, Xiaocan
%A Lin, Haoyu
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F li-lin-2026-fuzhoubench
%X Eastern Min Chinese (Fuzhou dialect, ISO 639-3 ‘cdo‘) has roughly ten million speakers yet no dedicated paper at any major NLP venue. Four properties make Fuzhou structurally distinct from Mandarin and challenging for current NLP: (1) a tone-sandhi system conditioned on the following syllable’s citation tone, (2) initial-consonant assimilation in non-final positions, (3) systematic literary-colloquial diglossia (45% of frequent characters carry \geq 2 distinct segmental readings), and (4) substantial lexical divergence (everyday concepts use entirely different morphemes from Mandarin). We release **FuzhouBench**, the first public NLP benchmark for Eastern Min, built on a community-maintained open dictionary: 22,890 word entries with 51,505 standardized Yngping romanizations, 11,276 explicit tone-sandhi annotations, 5,160 explanations, and 3.06 hours of paired dictionary-style audio from two native speakers. We define three tasks (grapheme-to-phoneme, sandhi prediction, definition retrieval) with full baselines. Frontier LLMs score under 3% word-exact on few-shot Yngping G2P, while a LoRA fine-tune on the released pairs reaches 17.2%: learnable, but far from saturated. A symbolic sandhi rule learned from the training pool beats every LLM by up to 18 points; and BM25 collapses from MRR 0.799 to 0.372 when queries shift from dictionary-extracted to colloquial Mandarin, partially recoverable via a two-stage hybrid + cross-encoder pipeline (MRR 0.716). Data, evaluation splits, and code are available at https://github.com/Nang-muoi-sing/FuzhouBench.
%U https://aclanthology.org/2026.emnlp-main.618/
%P 14194-14208

Markdown (Informal)

[FuzhouBench: A Benchmark for Eastern Min Chinese with Tone Sandhi Annotations](https://aclanthology.org/2026.emnlp-main.618/) (Li & Lin, EMNLP 2026)

• FuzhouBench: A Benchmark for Eastern Min Chinese with Tone Sandhi Annotations (Li & Lin, EMNLP 2026)

ACL

• Xiaocan Li and Haoyu Lin. 2026. FuzhouBench: A Benchmark for Eastern Min Chinese with Tone Sandhi Annotations. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 14194–14208, Budapest, Hungary. Association for Computational Linguistics.