<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="singhal-etal-2026-tokenization">
    <titleInfo>
        <title>Tokenization as Structural Inductive Bias: Grapheme-Constrained Subwords for Multilingual Modeling and Vocabulary Surgery</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Arush</namePart>
        <namePart type="family">Singhal</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jatin</namePart>
        <namePart type="family">Bedi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jasmeet</namePart>
        <namePart type="family">Singh</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>Subword tokenizers are optimized for compression, but their boundaries also define the prediction units seen by a language model. We introduce StructPiece, an unsupervised tokenizer that enforces Unicode Extended Grapheme Cluster (EGC) boundaries and biases merge induction toward corpus-induced affix patterns without external resources. Across nine languages and five scripts, StructPiece exposes a compression–representation frontier: it produces higher-fertility segmentations with worse bits-per-character but substantially lower token-level perplexity under matched architectures, vocabulary sizes, and token budgets. Fixed-budget, granularity-matched, and component ablations show that the effect is not explained by token count alone; structural vocabulary alignment is further tested through LLaMA-3.2-1B vocabulary surgery, replacing its 128k vocabulary with 32k StructPiece tokens via mean-pooled initialization. StructPiece recovers 89.4% of baseline zero-shot reasoning performance across morphologically rich languages (X-CSQA) after 2.6B adaptation tokens, compared with 76.4% recovery for naive BPE surgery under the same checkpoint. Furthermore, structurally aligned tokens produce higher mean-pooled embedding cosine coherence, suggesting better-conditioned vocabulary projections. These results suggest that token boundaries act as a controllable structural bias influencing prediction granularity and vocabulary adaptation. StructPiece is available at https://github.com/submissionemnlp26-beep/structpiece-code.</abstract>
    <identifier type="citekey">singhal-etal-2026-tokenization</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.972/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>22414</start>
            <end>22434</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Tokenization as Structural Inductive Bias: Grapheme-Constrained Subwords for Multilingual Modeling and Vocabulary Surgery
%A Singhal, Arush
%A Bedi, Jatin
%A Singh, Jasmeet
%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 singhal-etal-2026-tokenization
%X Subword tokenizers are optimized for compression, but their boundaries also define the prediction units seen by a language model. We introduce StructPiece, an unsupervised tokenizer that enforces Unicode Extended Grapheme Cluster (EGC) boundaries and biases merge induction toward corpus-induced affix patterns without external resources. Across nine languages and five scripts, StructPiece exposes a compression–representation frontier: it produces higher-fertility segmentations with worse bits-per-character but substantially lower token-level perplexity under matched architectures, vocabulary sizes, and token budgets. Fixed-budget, granularity-matched, and component ablations show that the effect is not explained by token count alone; structural vocabulary alignment is further tested through LLaMA-3.2-1B vocabulary surgery, replacing its 128k vocabulary with 32k StructPiece tokens via mean-pooled initialization. StructPiece recovers 89.4% of baseline zero-shot reasoning performance across morphologically rich languages (X-CSQA) after 2.6B adaptation tokens, compared with 76.4% recovery for naive BPE surgery under the same checkpoint. Furthermore, structurally aligned tokens produce higher mean-pooled embedding cosine coherence, suggesting better-conditioned vocabulary projections. These results suggest that token boundaries act as a controllable structural bias influencing prediction granularity and vocabulary adaptation. StructPiece is available at https://github.com/submissionemnlp26-beep/structpiece-code.
%U https://aclanthology.org/2026.emnlp-main.972/
%P 22414-22434

Markdown (Informal)

[Tokenization as Structural Inductive Bias: Grapheme-Constrained Subwords for Multilingual Modeling and Vocabulary Surgery](https://aclanthology.org/2026.emnlp-main.972/) (Singhal et al., EMNLP 2026)

• Tokenization as Structural Inductive Bias: Grapheme-Constrained Subwords for Multilingual Modeling and Vocabulary Surgery (Singhal et al., EMNLP 2026)

ACL

• Arush Singhal, Jatin Bedi, and Jasmeet Singh. 2026. Tokenization as Structural Inductive Bias: Grapheme-Constrained Subwords for Multilingual Modeling and Vocabulary Surgery. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 22414–22434, Budapest, Hungary. Association for Computational Linguistics.