Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 43

Padamitra: Grounded Glossary Generation for Classical Sanskrit Padamitra:古典梵语接地词汇表生成

Introduces "grounded glossary generation," a novel NLP task that formalizes the traditional Sanskrit patha commentary practice into a structured, evaluable objective requiring models to recover semantically meaningful phrases and produce translation-grounded meanings from sloka-translation pairs Constructs a large-scale benchmark of 31,316 sloka-translation-glossary triples drawn from the Valmiki Ramayana and Srimad Bhagavatam, paired with two evaluation metrics: Jaccard similarity for phrase re 提出"接地词汇表生成"新任务,将传统梵语patha注释实践形式化为可评估的NLP目标 构建31,316条sloka-translation-glossary三元组基准,源自《罗摩衍那》和《薄伽梵往世书》 设计双指标评估体系:Jaccard衡量短语恢复,Meaning Faithfulness衡量语义一致性 在Gemma-3n-E4B、Gemma-3-12B、Phi-4、Qwen3.5-9B上验证,指令微调显著优于提示工程 错误分析揭示sandhi连声和samasa复合词过度分割是主要瓶颈,形态建模是关键挑战

55
Hot 热度
72
Quality 质量
58
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces "grounded glossary generation," a novel NLP task that formalizes the traditional Sanskrit patha commentary practice into a structured, evaluable objective requiring models to recover semantically meaningful phrases and produce translation-grounded meanings from sloka-translation pairs
  • Constructs a large-scale benchmark of 31,316 sloka-translation-glossary triples drawn from the Valmiki Ramayana and Srimad Bhagavatam, paired with two evaluation metrics: Jaccard similarity for phrase recovery and Meaning Faithfulness for semantic consistency
  • Evaluates multiple modern LLMs (Gemma-3n-E4B, Gemma-3-12B, Phi-4, Qwen3.5-9B) across zero-shot, few-shot, and instruction fine-tuned settings, finding that instruction fine-tuning substantially outperforms prompting strategies
  • Identifies over-segmentation of sandhi and samasa compounds as the dominant failure mode, highlighting morphological modeling as the key bottleneck for faithful Sanskrit lexical decomposition

Why It Matters

This work bridges classical Sanskrit scholarship and modern NLP by formalizing a traditional commentary practice into a rigorous benchmark, enabling systematic evaluation of LLMs on low-resource, morphologically rich languages. It demonstrates that even state-of-the-art models struggle with complex morphological phenomena like sandhi and samasa, pointing to a critical gap in current architectures for handling agglutinative and compounding languages.

Technical Details

  • Task formulation: Grounded glossary generation requires models to decompose Sanskrit slokas into meaningful lexical units and assign translation-grounded meanings, mirroring the patha commentary tradition used by classical Sanskrit scholars
  • Benchmark construction: 31,316 triples from two major Sanskrit texts (Valmiki Ramayana and Srimad Bhagavatam), with evaluation via Jaccard score for phrase-level recovery accuracy and Meaning Faithfulness metric for semantic consistency against reference glossaries
  • Model evaluation: Tested Gemma-3n-E4B, Gemma-3-12B, Phi-4, and Qwen3.5-9B across three settings—zero-shot prompting, few-shot prompting, and instruction fine-tuning—with explicit segmentation strategies yielding additional performance gains
  • Error analysis: Over-segmentation of sandhi (phonological concatenation) and samasa (compound) forms is the primary failure mode, indicating that current LLMs lack robust morphological decomposition capabilities for highly inflected languages

Industry Insight

  • Morphological modeling remains a critical bottleneck for deploying LLMs on low-resource, morphologically rich languages; investing in explicit morphological preprocessing or specialized tokenization could unlock significant performance gains
  • Instruction fine-tuning on domain-specific, culturally grounded tasks substantially outperforms general prompting, suggesting that targeted fine-tuning on classical text corpora is a viable strategy for building specialized language models
  • The formalization of traditional scholarly practices (like patha commentary) into evaluable NLP benchmarks offers a replicable template for bringing other heritage knowledge systems into modern AI evaluation frameworks

TL;DR

  • 提出"接地词汇表生成"新任务,将传统梵语patha注释实践形式化为可评估的NLP目标
  • 构建31,316条sloka-translation-glossary三元组基准,源自《罗摩衍那》和《薄伽梵往世书》
  • 设计双指标评估体系:Jaccard衡量短语恢复,Meaning Faithfulness衡量语义一致性
  • 在Gemma-3n-E4B、Gemma-3-12B、Phi-4、Qwen3.5-9B上验证,指令微调显著优于提示工程
  • 错误分析揭示sandhi连声和samasa复合词过度分割是主要瓶颈,形态建模是关键挑战

为什么值得看

该研究首次将古典梵语文献的注释传统转化为可量化的NLP任务,为低资源语言处理提供新范式。通过构建大规模梵语基准并系统评估主流开源模型,揭示了形态复杂语言在语义分解任务中的核心难点。

技术解析

  • 任务定义:grounded glossary generation要求模型从梵语诗节(sloka)与翻译对中,恢复语义有意义的短语并生成翻译接地含义,形式化传统patha注释实践
  • 数据集规模:31,316个sloka-translation-glossary三元组,来源于Valmiki Ramayana和Srimad Bhagavatam两大经典梵语文本
  • 评估指标:Jaccard相似度用于短语恢复准确性评估,Meaning Faithfulness用于语义一致性评估
  • 模型实验:测试Gemma-3n-E4B、Gemma-3-12B、Phi-4、Qwen3.5-9B的zero-shot、few-shot和instruction fine-tuned变体
  • 关键发现:指令微调相比提示工程带来显著提升,显式分割策略进一步改善性能;sandhi连声和samasa复合词的过度分割是主导性错误模式

行业启示

  • 低资源/古典语言处理可借鉴"传统学术实践形式化"路径,将人文领域知识转化为可计算NLP任务
  • 形态丰富语言的处理瓶颈在于复杂词形分析,未来需加强morphological modeling能力建设
  • 开源小模型(4B-12B)在特定领域任务经指令微调后可达到实用水平,为垂直领域部署提供可行方案

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 Benchmark 基准测试 Dataset 数据集 Evaluation 评测 LLM 大模型