<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="syvokon-2026-dictionary">
    <titleInfo>
        <title>Dictionary-Based Speculative Decoding for Non-Latin-Script Languages</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Oleksiy</namePart>
        <namePart type="family">Syvokon</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-05</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the Fifth Ukrainian Natural Language Processing Conference (UNLP 2026)</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Mariana</namePart>
            <namePart type="family">Romanyshyn</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Lviv, Ukraine</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-359-3</identifier>
    </relatedItem>
    <abstract>Large language models tokenize non-Latin-script languagesinefficiently: a single word in Ukrainian or Crimean Tatar is split intotwo to three times as many tokens as its English equivalent. We propose_dictionary-based speculative decoding_ (DictSpec), which acceleratesinference by proposing draft continuations from a static n-gram lookuptable built offline from an unlabeled corpus. The lookup table requiresno trainable parameters or GPU resources, is inexpensive to construct,adds under 5 MB of memory overhead, and can be reused across modelsthat share a tokenizer. We evaluate DictSpec on Ukrainian and Crimean Tatar(Cyrillic and Latin scripts), implementing a vLLM plugin to benchmarkfive models ranging from 3B to 70B parameters on consumer- andserver-grade GPUs. In controlled emulation, DictSpec reduces verificationsteps by up to 1.65×, with gains correlating substantially with tokenizerfertility. In live vLLM serving, pure DictSpec gives modest speedups,while a hybrid with prompt-local n-gram speculation reaches up to 1.76×.We release our code and vLLM plugin as opensource.</abstract>
    <identifier type="citekey">syvokon-2026-dictionary</identifier>
    <location>
        <url>https://aclanthology.org/2026.unlp-1.15/</url>
    </location>
    <part>
        <date>2026-05</date>
        <extent unit="page">
            <start>169</start>
            <end>183</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Dictionary-Based Speculative Decoding for Non-Latin-Script Languages
%A Syvokon, Oleksiy
%Y Romanyshyn, Mariana
%S Proceedings of the Fifth Ukrainian Natural Language Processing Conference (UNLP 2026)
%D 2026
%8 May
%I Association for Computational Linguistics
%C Lviv, Ukraine
%@ 979-8-89176-359-3
%F syvokon-2026-dictionary
%X Large language models tokenize non-Latin-script languagesinefficiently: a single word in Ukrainian or Crimean Tatar is split intotwo to three times as many tokens as its English equivalent. We propose_dictionary-based speculative decoding_ (DictSpec), which acceleratesinference by proposing draft continuations from a static n-gram lookuptable built offline from an unlabeled corpus. The lookup table requiresno trainable parameters or GPU resources, is inexpensive to construct,adds under 5 MB of memory overhead, and can be reused across modelsthat share a tokenizer. We evaluate DictSpec on Ukrainian and Crimean Tatar(Cyrillic and Latin scripts), implementing a vLLM plugin to benchmarkfive models ranging from 3B to 70B parameters on consumer- andserver-grade GPUs. In controlled emulation, DictSpec reduces verificationsteps by up to 1.65×, with gains correlating substantially with tokenizerfertility. In live vLLM serving, pure DictSpec gives modest speedups,while a hybrid with prompt-local n-gram speculation reaches up to 1.76×.We release our code and vLLM plugin as opensource.
%U https://aclanthology.org/2026.unlp-1.15/
%P 169-183

Markdown (Informal)

[Dictionary-Based Speculative Decoding for Non-Latin-Script Languages](https://aclanthology.org/2026.unlp-1.15/) (Syvokon, UNLP 2026)

• Dictionary-Based Speculative Decoding for Non-Latin-Script Languages (Syvokon, UNLP 2026)

ACL

• Oleksiy Syvokon. 2026. Dictionary-Based Speculative Decoding for Non-Latin-Script Languages. In Proceedings of the Fifth Ukrainian Natural Language Processing Conference (UNLP 2026), pages 169–183, Lviv, Ukraine. Association for Computational Linguistics.