Research Papers 论文研究 18h ago Updated 2h ago 更新于 2小时前 35

What Counts as a Mistake? Annotating Recitation Events in Quran Memorization Transcripts What Counts as a Mistake? Annotating Recitation Events in Quran Memorization Transcripts

Human-annotated dataset of 100 Quran recitation recordings with 162 localized error events across ten annotation labels, addressing the challenge of distinguishing mistakes from repetitions, repairs, and accepted spelling variations Executable evaluator jointly scores label correctness and word-position localization, achieving label-aware F1 of 0.525 and localization F1 of 0.826 on a plain diff baseline Pilot evaluation across three coding agents and eight models shows wide performance variance 提出古兰经诵读转录中错误标注的系统方法,需区分未解决错误、重复、修复、开场公式和可接受拼写差异 完成100个生产录音案例的人工标注,包含348个评分单元和162个本地化事件,使用十个组合标签 开发可执行评估器同时评分标签和词位置,简单diff达到标签感知F1 0.525和定位F1 0.826 初步试点中八个单次20分钟运行跨越三个编码代理和八个模型,标签感知F1从0.143到0.892不等 发现970/972个黄金事件实例获得重叠预测,剩余问题在于跨度范围和标签边界的惯例而非检测能力

50
Hot 热度
50
Quality 质量
50
Impact 影响力

Analysis 深度分析

TL;DR

  • Human-annotated dataset of 100 Quran recitation recordings with 162 localized error events across ten annotation labels, addressing the challenge of distinguishing mistakes from repetitions, repairs, and accepted spelling variations
  • Executable evaluator jointly scores label correctness and word-position localization, achieving label-aware F1 of 0.525 and localization F1 of 0.826 on a plain diff baseline
  • Pilot evaluation across three coding agents and eight models shows wide performance variance (label-aware F1 0.143 to 0.892), with seven of eight runs significantly outperforming baselines
  • Core finding: detection is largely solved; the remaining challenge is conventional—span boundaries and label assignments are stipulated by adjudication rather than visible in the text
  • Seven of 162 events defeat all six same-day runs, five due to a single orthographic rule, revealing that annotation interfaces must be validated before interpreting baseline failures

Why It Matters

This work addresses a critical gap in domain-specific ASR evaluation: how to systematically annotate and measure error types in specialized recitation tasks where accepted variations (spelling, repetition, repairs) must be distinguished from genuine mistakes. For AI practitioners building evaluation pipelines for niche linguistic domains, it demonstrates that baseline failures may reflect annotation interface issues rather than algorithmic limitations, and that what appears to be a detection problem may actually be a convention problem requiring human adjudication.

Technical Details

  • Dataset: 100 production Quran recitation recordings, human-annotated with 348 scored units and 162 localized events across ten combined labels (mistakes, repetitions, repairs, opening formulas, accepted spelling differences)
  • Evaluator: Executable joint scorer that evaluates both label correctness and word-position localization simultaneously, reporting label-aware F1, localization F1, and exact-span F1 metrics
  • Baselines: Plain diff achieves label-aware F1 0.525 and localization F1 0.826; adapted production cleaner/alignment components reach 0.518 and 0.786 respectively, with exact-span F1 0.505 for both
  • Pilot: Eight 20-minute runs across three coding agents and eight models, with label-aware F1 ranging from 0.143 to 0.892; seven runs significantly outperform baselines while one collapses below naive diff due to missing normalization
  • Key insight: 970 of 972 gold-event instances draw overlapping predictions across runs, indicating detection is solved; the remaining 972-970=2 discrepancy and all span-boundary disagreements are conventional, not algorithmic

Industry Insight

  • Annotation validation is prerequisite: Before interpreting baseline failures in specialized NLP tasks, validate that annotation interfaces correctly capture the ground truth—five of the seven unbeatable events were defeated by a single orthographic rule, suggesting interface issues rather than model limitations
  • Convention over detection: In domain-specific evaluation tasks, once detection capability reaches a threshold, further gains require resolving conventional disagreements (span boundaries, label assignments) through human adjudication, not algorithmic improvements
  • Pilot design matters: The wide performance variance (0.143 to 0.892 F1) across eight runs demonstrates that even with the same task definition, implementation choices (normalization steps, alignment components) dramatically affect results—standardized pilot protocols are essential for fair comparison

TL;DR

  • 提出古兰经诵读转录中错误标注的系统方法,需区分未解决错误、重复、修复、开场公式和可接受拼写差异
  • 完成100个生产录音案例的人工标注,包含348个评分单元和162个本地化事件,使用十个组合标签
  • 开发可执行评估器同时评分标签和词位置,简单diff达到标签感知F1 0.525和定位F1 0.826
  • 初步试点中八个单次20分钟运行跨越三个编码代理和八个模型,标签感知F1从0.143到0.892不等
  • 发现970/972个黄金事件实例获得重叠预测,剩余问题在于跨度范围和标签边界的惯例而非检测能力

为什么值得看

这篇论文为宗教文本的ASR转录错误标注提供了系统方法,对阿拉伯语语音识别和宗教文本处理具有重要参考价值。研究揭示了算法性能已接近人类水平,主要挑战在于标注惯例和边界定义而非检测能力。

技术解析

  • 标注数据集:100个生产录音案例,348个评分单元,162个本地化事件,十个组合标签
  • 评估方法:可执行评估器同时评分标签和词位置,简单diff达到标签感知F1 0.525和定位F1 0.826
  • 基线结果:适应的生产清理器/对齐组件达到0.518和0.786,精确跨度F1为0.505
  • 试点实验:八个单次20分钟运行,三个编码代理,八个模型,标签感知F1从0.143到0.892不等
  • 关键发现:七个事件击败所有六个同一天运行的模型,其中五个涉及一个正字法规则

行业启示

  • 宗教文本的ASR转录需要专门的标注方法和评估体系,不能简单套用通用NLP方法
  • 算法性能已接近人类水平,主要挑战在于标注惯例和边界定义而非检测能力
  • 建议在构建系统前先进行标注,以确保评估的准确性和可比性

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。