Dataset Card for BanglaReDup

BanglaReDup is the first publicly available 20,000-row Bangla dataset explicitly annotated to distinguish Repetition Disfluency, Morphological Reduplication, and Neither in noisy Automatic Speech Recognition (ASR) transcripts. The data comes from real YouTube Bangla ASR subtitles and includes both coarse labels and fine-grained reduplication subtypes.

Dataset Details

Dataset Description

BanglaReDup addresses a major linguistic challenge in Bangla ASR transcripts: identical word-word repetitions can represent either grammatically meaningful reduplication or unintentional disfluency. Standard disfluency-correction systems remove both, which harms meaning. This dataset enables accurate classification of such cases.

• Curated by: Zaara Zabeen Arpa, Sadnam Sakib Apurbo, Nazia Karim Khan Oishee, Ajwad Abrar
• Funded by: Islamic University of Technology (IUT), Bangladesh
• Shared by: Ajwad Abrar
• Language(s): Bangla (bn)
• License: CC BY-NC-SA 4.0

Dataset Sources

• Repository: https://huggingface.co/datasets/ajwad-abrar/BanglaReDup
• Paper: Distinguishing Repetition Disfluency from Morphological Reduplication in Bangla ASR Transcripts (AI & Society, Springer Nature)
• Demo: Not available yet

---

Uses

Direct Use

• Training classification models for reduplication vs disfluency
• Building semantic-preserving Bangla ASR post-processing systems
• Research on Bangla linguistics and low-resource NLP
• Evaluating LLM prompting vs fine-tuning

Out-of-Scope Use

• Not intended for full ASR model training
• Not suitable for sentiment, summarization, translation, or demographic inference
• Not appropriate for applications requiring speaker identity

---

Dataset Structure

Each data sample includes:

• context — 10 tokens before and after the repeated word
• repetition_phrase — the target repeated segment
• label — one of: Reduplication, Repetition, Neither
• fine_grained_reduplication — one of nine subtypes (only for Reduplication)
• bio_tags — B/I tags for the repeated words

Splits:

• 20,000 total rows
• 335-row test set (used in the paper)
• Train-validation recommended split: 80/20

---

Dataset Creation

Curation Rationale

Bangla ASR transcripts frequently misinterpret reduplication as disfluency. No annotated Bangla resource existed to resolve this ambiguity, motivating the creation of BanglaReDup.

Source Data

Data Collection and Processing

• Scraped Bangla educational YouTube channels using yt-dlp
• Only videos with Bangla ASR subtitles (bn-orig) were selected
• Transcripts were cleaned, tokenized, and filtered
• Consecutive identical tokens extracted with 10-word context windows
• BIO tagging applied for structural clarity

Who are the source data producers?

• Native Bangla speakers in public YouTube educational videos
• YouTube’s ASR system generated the raw transcripts

---

Annotations

Annotation Process

A two-stage hybrid process:

1. LLM-assisted initial annotation using GPT-4o with few-shot prompts
2. Manual expert verification by native annotators
3. Fine-grained subcategorization into nine reduplication types, assisted by LLM and verified manually

Who are the annotators?

• Three native Bangla-speaking annotators and the corresponding author (supervision)

Personal and Sensitive Information

• No personal identity or sensitive data
• All text is from publicly accessible educational content
• Data anonymized and stripped of metadata

---

Bias, Risks, and Limitations

• Class imbalance: Reduplication (66.3%) dominates the dataset
• Domain-specific: Only educational YouTube content included
• ASR noise: Some contextual tokens may contain transcription errors
• LLM-initial bias: Initial model predictions may influence ambiguous cases

Recommendations

• Users should consider rebalancing when training models
• Models should be tested on conversational or cross-domain corpora before deployment
• Avoid demographic or speaker-based inference tasks

---

Citation

BibTeX

  @misc{arpa2025distinguishingrepetitiondisfluencymorphological,
      title={Distinguishing Repetition Disfluency from Morphological Reduplication in Bangla ASR Transcripts: A Novel Corpus and Benchmarking Analysis}, 
      author={Zaara Zabeen Arpa and Sadnam Sakib Apurbo and Nazia Karim Khan Oishee and Ajwad Abrar},
      year={2025},
      eprint={2511.13159},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.13159}, 
}