Research Papers 论文研究 5h ago Updated 53m ago 更新于 53分钟前 45

Beyond Two Bytes per Letter: Tokenization Overhead in Cyrillic AI Systems 超越每字母两字节:西里尔字母AI系统中的分词开销

Ukrainian and other underrepresented Cyrillic-script languages experience 68-121% token overhead on modern tokenizers and 220% on older cl100k tokenizers compared to English Training data allocation is a primary contributor to Cyrillic tokenization overhead, with efficiency favoring the script more prevalent in web data LLMLingua-2 reduces Ukrainian input length by 47-49% on an e-commerce RAG benchmark with no compression-induced value losses A balanced byte-level BPE tokenizer with a 200K vocab 现代多语言分词器对乌克兰语等西里尔字母语言的碎片化程度显著高于英语,造成68-121%的分词开销(旧版cl100k高达220%) 研究覆盖9个生产级分词器、5种语言,共分析837万词形,发现分词效率与训练数据中西里尔字母词汇分配呈负相关 两种缓解策略有效:LLMLingua-2在电商RAG基准上减少47-49%输入长度且无价值损失;平衡字节级BPE分词器将UK/EN比率从2.22x降至1.30x 罗马化策略反而增加2-19%分词数量,训练数据分配是造成西里尔字母分词开销的根本原因

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Ukrainian and other underrepresented Cyrillic-script languages experience 68-121% token overhead on modern tokenizers and 220% on older cl100k tokenizers compared to English
  • Training data allocation is a primary contributor to Cyrillic tokenization overhead, with efficiency favoring the script more prevalent in web data
  • LLMLingua-2 reduces Ukrainian input length by 47-49% on an e-commerce RAG benchmark with no compression-induced value losses
  • A balanced byte-level BPE tokenizer with a 200K vocabulary cap reduces the UK/EN token ratio from 2.22x to 1.30x
  • Romanization increases Ukrainian token counts by 2-19% on most tokenizers, making it an ineffective mitigation strategy

Why It Matters

This research directly impacts cost and performance for AI systems serving Cyrillic-speaking users, particularly in regions like Ukraine where multilingual AI adoption is growing. The quantified overhead reveals a systemic inequity in how modern tokenizers handle underrepresented languages, which translates to higher inference costs and reduced context capacity for non-English applications. For practitioners building multilingual systems, these findings provide both a diagnostic framework and actionable mitigation strategies.

Technical Details

  • Evaluated nine production tokenizers across five languages with standardized Cyrillic and Latin representations, covering 8.37 million word forms using the BrUK and Brown corpora for benchmarking
  • Measured token overhead through full-text fertility analysis, establishing Ukrainian at 68-121% overhead on modern tokenizers and 220% on cl100k
  • Tested LLMLingua-2 on an e-commerce RAG benchmark comprising 1,536 products and 145 queries, achieving 47-49% input reduction with zero value loss across 80 retrievable cases
  • Designed a balanced byte-level BPE tokenizer with a 200K vocabulary cap (converging at 158,184 entries), reducing the held-out UK/EN ratio from 2.22x to 1.30x
  • Found a negative but non-significant association between Cyrillic vocabulary allocation and overhead (Spearman rho = -0.536, p = 0.215, n = 7), suggesting data allocation matters but requires larger samples for statistical confidence

Industry Insight

  • Organizations deploying multilingual AI should audit their tokenizer choices for Cyrillic-language support, as the 2-3x overhead directly impacts token costs and effective context window utilization
  • Investing in balanced tokenizer training with explicit vocabulary caps for underrepresented scripts can yield significant efficiency gains without sacrificing English performance
  • LLMLingua-2 presents a practical, drop-in mitigation for existing deployments, but long-term solutions require rethinking tokenizer training data allocation rather than relying on romanization or compression alone

TL;DR

  • 现代多语言分词器对乌克兰语等西里尔字母语言的碎片化程度显著高于英语,造成68-121%的分词开销(旧版cl100k高达220%)
  • 研究覆盖9个生产级分词器、5种语言,共分析837万词形,发现分词效率与训练数据中西里尔字母词汇分配呈负相关
  • 两种缓解策略有效:LLMLingua-2在电商RAG基准上减少47-49%输入长度且无价值损失;平衡字节级BPE分词器将UK/EN比率从2.22x降至1.30x
  • 罗马化策略反而增加2-19%分词数量,训练数据分配是造成西里尔字母分词开销的根本原因

为什么值得看

本文首次系统量化了西里尔字母语言在现代AI分词器中的效率劣势,揭示了多语言AI系统中的隐性不平等。研究提出的缓解方案为开发者和企业提供了可落地的优化路径,对降低多语言AI系统成本、提升上下文容量具有直接指导意义。

技术解析

  • 实验规模:覆盖9个生产级分词器、5种语言(含标准化西里尔与拉丁表示),分析837万词形,基于BrUK和Brown语料库进行全文生育率测量
  • 开销量化:乌克兰语在现代分词器上产生68-121%分词开销,旧版cl100k高达220%;Spearman相关系数-0.536(p=0.215)显示西里尔词汇分配与开销负相关
  • LLMLingua-2方案:在1,536产品、145查询的电商RAG基准上,将乌克兰语输入长度压缩47-49%,80个可检索案例中无压缩导致的价值损失
  • 平衡BPE分词器:采用200K词表上限训练,实际收敛于158,184条目,将保留集UK/EN比率从2.22x降至1.30x
  • 罗马化实验:在大多数分词器上,罗马化反而使乌克兰语分词数量增加2-19%,证明转换脚本无法解决根本问题

行业启示

  • 多语言公平性需纳入AI系统设计:当前分词器对西里尔字母语言的系统性劣势表明,多语言AI存在隐性成本不平等,开发者应在模型选型时评估目标语言的token效率
  • 训练数据分配决定分词效率:研究证实web数据中脚本 prevalence 直接影响分词效率,建议企业在构建多语言语料库时主动平衡西里尔字母数据比例
  • 缓解策略可分层部署:推理阶段可采用LLMLingua-2等压缩技术降低成本,tokenizer设计阶段可通过平衡训练数据优化效率,两者结合可实现显著改善

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Research 科学研究 Dataset 数据集 Benchmark 基准测试 Evaluation 评测