Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PersianPunc: A Large-Scale Dataset and BERT-Based Approach for Persian Punctuation Restoration

Mohammad Javad Ranjbar Kalahroodi, Heshaam Faili, Azadeh Shakery

Abstract

Punctuation restoration is essential for improving the readability and downstream utility of automatic speech recognition (ASR) outputs, yet remains underexplored for Persian despite its importance. We introduce PersianPunc, a large-scale, high-quality dataset of 17 million samples for Persian punctuation restoration, constructed through systematic aggregation and filtering of existing textual resources. We formulate punctuation restoration as a token-level sequence labeling task and fine-tune ParsBERT to achieve strong performance. Through comparative evaluation, we demonstrate that while large language models can perform punctuation restoration, they suffer from critical limitations: over-correction tendencies that introduce undesired edits beyond punctuation insertion (particularly problematic for speech-to-text pipelines) and substantially higher computational requirements. Our lightweight BERT-based approach achieves a macro-averaged F1 score of 91.33% on our test set while maintaining efficiency suitable for real-time applications. We make our dataset (https://huggingface.co/datasets/MohammadJRanjbar/persian-punctuation-restoration) and model (https://huggingface.co/MohammadJRanjbar/parsbert-persian-punctuation) publicly available to facilitate future research in Persian NLP and provide a scalable framework applicable to other morphologically rich, low-resource languages.

PersianPunc: A Large-Scale Dataset and BERT-Based Approach for Persian Punctuation Restoration

Abstract

Punctuation restoration is essential for improving the readability and downstream utility of automatic speech recognition (ASR) outputs, yet remains underexplored for Persian despite its importance. We introduce PersianPunc, a large-scale, high-quality dataset of 17 million samples for Persian punctuation restoration, constructed through systematic aggregation and filtering of existing textual resources. We formulate punctuation restoration as a token-level sequence labeling task and fine-tune ParsBERT to achieve strong performance. Through comparative evaluation, we demonstrate that while large language models can perform punctuation restoration, they suffer from critical limitations: over-correction tendencies that introduce undesired edits beyond punctuation insertion (particularly problematic for speech-to-text pipelines) and substantially higher computational requirements. Our lightweight BERT-based approach achieves a macro-averaged F1 score of 91.33% on our test set while maintaining efficiency suitable for real-time applications. We make our dataset (https://huggingface.co/datasets/MohammadJRanjbar/persian-punctuation-restoration) and model (https://huggingface.co/MohammadJRanjbar/parsbert-persian-punctuation) publicly available to facilitate future research in Persian NLP and provide a scalable framework applicable to other morphologically rich, low-resource languages.
Paper Structure (32 sections, 2 figures, 6 tables)

This paper contains 32 sections, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Punctuation Restoration as a Sequence Modeling Task
  4. Transformer-Based Approaches for Punctuation Restoration
  5. State of Persian Text Processing and Punctuation
  6. Methodology
  7. Dataset Construction
  8. Data Sources and Collection Strategy
  9. Preprocessing and Quality Control
  10. Normalization Pipeline
  11. Sentence Segmentation and Filtering
  12. Rationale for Filtering Criteria
  13. Deduplication and Dataset Splitting
  14. Dataset Statistics and Punctuation Analysis
  15. Punctuation Distribution
  16. ...and 17 more sections

Figures (2)

  • Figure 1: Persian punctuation restoration dramatically affects semantic interpretation. Minimal punctuation changes transform sentence meaning from negative to positive sentiment.
  • Figure 2: Prompt used for zero-shot evaluation of GPT-4o and GPT-4o-mini on Persian punctuation restoration. The system is explicitly instructed to only add punctuation marks without altering the original text in any way.