Research Papers 论文研究 3h ago Updated 1h ago 更新于 1小时前 48

BHARATI: Morphology-Aware Tokenizers for Classical Indian Languages with Subword Fertility Analysis BHARATI:面向古典印度语言的形态感知分词器与子词生育力分析

BHARATI introduces morphology-aware SentencePiece BPE tokenizers specifically designed for classical Indian languages (Sanskrit, Tamil, Telugu, Kannada, Malayalam) alongside English and Hindi. The v3 tokenizer achieves a subword fertility of 2.6 tokens per Indian Knowledge System technical term, significantly outperforming GPT-2's 5.25 and the mBART-50 baseline's 3.75. On a held-out IKS test set, BHARATI v3 reduces sequence length by approximately 90% compared to byte-level encodings and roughly 提出BHARATI,一种针对古典印度语言(如梵语、泰米尔语等)的形态感知分词器,解决标准子词算法在处理黏着性形态和音韵融合时的低效问题。 基于781MB平衡语料训练出三个版本的SentencePiece BPE分词器(v1-v3),最终版本支持七种语言的原生脚本覆盖,词汇量为32,000。 在印度知识系统(IKS)技术术语上,v3平均每个术语仅需2.6个token,显著优于GPT-2(5.25)和多语言基线(3.75)。 测试集显示v3使序列长度比GPT-2减少约90%,比mBART-50基线减少约25%,直接提升下游模型的有效上下文长度。 开源提供分词器模型、训练脚本及评估基准,促进古典印度

65
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • BHARATI introduces morphology-aware SentencePiece BPE tokenizers specifically designed for classical Indian languages (Sanskrit, Tamil, Telugu, Kannada, Malayalam) alongside English and Hindi.
  • The v3 tokenizer achieves a subword fertility of 2.6 tokens per Indian Knowledge System technical term, significantly outperforming GPT-2's 5.25 and the mBART-50 baseline's 3.75.
  • On a held-out IKS test set, BHARATI v3 reduces sequence length by approximately 90% compared to byte-level encodings and roughly 25% relative to the mBART-50 multilingual baseline.
  • The system supports native script coverage across all seven target languages within a balanced 781 MB corpus, addressing inefficiencies caused by agglutinative morphology and productive sandhi in standard tokenizers.

Why It Matters

This work addresses a critical bottleneck in deploying Large Language Models for low-resource and morphologically complex Indian languages. By optimizing tokenization for specific linguistic features like sandhi and domain-specific vocabularies, BHARATI directly increases effective context length and computational efficiency, which is essential for building capable AI systems that preserve and process India's vast digital heritage and knowledge systems.

Technical Details

  • Tokenizer Architecture: Utilizes SentencePiece Byte-Pair Encoding (BPE) trained on a curated, balanced corpus rather than relying on general-purpose web data.
  • Iterative Development: Three versions were developed: v1 focused on English/Sanskrit with fallback for Tamil; v2 expanded to four languages; v3 achieved full native subword coverage for all seven languages (English, Hindi, Sanskrit, Tamil, Telugu, Kannada, Malayalam).
  • Vocabulary Size: The final model employs a vocabulary size of 32,000 tokens.
  • Evaluation Metrics: Performance measured via subword fertility (tokens per technical term) and sequence length reduction on a dataset of 490 IKS-domain sentences (70 per language).
  • Linguistic Handling: Specifically engineered to handle agglutinative morphology and phonological fusion (sandhi) which typically cause standard tokenizers to fragment words excessively.

Industry Insight

  • Efficiency Gains: Reducing sequence length by 90% relative to byte-level encoding allows for significantly longer context windows or reduced training costs for downstream models processing Indic texts.
  • Domain Specialization: The success of reserving specific terms as single tokens suggests that hybrid approaches combining general BPE with domain-specific vocabulary injection are highly effective for specialized fields like Indian Knowledge Systems.
  • Open Source Reproducibility: The release of training scripts and evaluation benchmarks provides a replicable framework for developing similar morphology-aware tokenizers for other under-resourced or morphologically complex language families globally.

TL;DR

  • 提出BHARATI,一种针对古典印度语言(如梵语、泰米尔语等)的形态感知分词器,解决标准子词算法在处理黏着性形态和音韵融合时的低效问题。
  • 基于781MB平衡语料训练出三个版本的SentencePiece BPE分词器(v1-v3),最终版本支持七种语言的原生脚本覆盖,词汇量为32,000。
  • 在印度知识系统(IKS)技术术语上,v3平均每个术语仅需2.6个token,显著优于GPT-2(5.25)和多语言基线(3.75)。
  • 测试集显示v3使序列长度比GPT-2减少约90%,比mBART-50基线减少约25%,直接提升下游模型的有效上下文长度。
  • 开源提供分词器模型、训练脚本及评估基准,促进古典印度语言的NLP研究与应用。

为什么值得看

该工作填补了古典印度语言在自然语言处理中的关键空白,通过专门设计的分词器显著提升这些语言在AI系统中的表示效率与性能。对于致力于多语言建模、文化遗产数字化或区域语言AI落地的从业者而言,其方法论与开源资源具有直接参考价值。

技术解析

  • 采用SentencePiece BPE算法构建分词器,针对古典印度语言的黏着性形态和sandhi现象进行优化,避免通用分词器导致的过度切分。
  • 训练数据为781MB平衡语料,涵盖英语、印地语、梵语、泰米尔语、泰卢固语、卡纳达语和马拉雅拉姆语七种语言,确保跨语言代表性。
  • 分词器分三阶段演进:v1仅支持英/梵语且泰米尔语依赖字节回退;v2扩展至四种语言并改进南部语言字节级fallback;v3实现七种语言原生子词全覆盖。
  • 引入“子词生育率”分析指标,量化不同分词器对IKS术语的处理效率,证明v3在保留语义完整性的同时大幅压缩token数量。
  • 在490句领域测试集(每语言70句)中验证效果,结合开放评估脚本复现结果,强调实际部署中的上下文长度增益。

行业启示

  • 推动AI领域对非主流语言特别是古典语言的重视,鼓励开发定制化分词策略以应对特殊语言结构,而非简单套用通用方案。
  • 为多语言大模型的本地化适配提供可复用范式,尤其在知识密集型场景(如宗教文本、历史文献处理)中展现显著优势。
  • 呼吁建立更多面向特定文化语境的开源工具链,加速全球语言多样性在人工智能时代的保护与创新应用。

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。

Research 科学研究 Open Source 开源 LLM 大模型