--- language: - en - bn - gu - hi - mr - ur - kn - ta - te pretty_name: "STAR-IL: A Dataset for Style-Aware Machine Translation of Product Reviews in Indian Languages" tags: - machine-translation - style-transfer - indian-languages - product-reviews - e-commerce - parallel-corpus license: cc-by-4.0 task_categories: - translation size_categories: - 10K Language Family Electronics Fashion Indo-Aryan Dravidian Indo-Aryan Dravidian Language Ben Guj Hin Mar Urd Kan Tam Tel Ben Guj Hin Mar Urd Kan Tam Tel Samples 4391 4491 4486 4489 4471 4488 4490 4491 2400 3199 2199 2200 3963 2199 1697 2200 Sample Length 47.86 56.55 15.93 14.23 16.52 55.13 67.15 61.97 259.79 310.23 80.98 72.65 87.90 307.15 334.91 355.34 TTR% 0.26 0.32 6.89 9.45 5.72 0.44 0.32 0.25 0.10 0.12 3.96 0.09 1.82 0.37 0.07 0.10 ### Overall Summary by Language | Language | ISO Code | Family | Electronics Samples | Fashion Samples | Total Samples | | :--- | :---: | :--- | :---: | :---: | :---: | | **Bengali** (`Ben`) | `bn` | Indo-Aryan | 4,391 | 2,400 | **6,791** | | **Gujarati** (`Guj`) | `gu` | Indo-Aryan | 4,491 | 3,199 | **7,690** | | **Hindi** (`Hin`) | `hi` | Indo-Aryan | 4,486 | 2,199 | **6,685** | | **Marathi** (`Mar`) | `mr` | Indo-Aryan | 4,489 | 2,200 | **6,689** | | **Urdu** (`Urd`) | `ur` | Indo-Aryan | 4,471 | 3,963 | **8,434** | | **Kannada** (`Kan`) | `kn` | Dravidian | 4,488 | 2,199 | **6,687** | | **Tamil** (`Tam`) | `ta` | Dravidian | 4,490 | 1,697 | **6,187** | | **Telugu** (`Tel`) | `te` | Dravidian | 4,491 | 2,200 | **6,691** | | **Total** | — | — | **35,797** | **20,057** | **55,854** | ## Repository Structure The dataset is organized into distinct directories corresponding to the source data, translated data, and annotation guidelines: ``` STAR-IL-Corpus/ ├── guidelines/ │ └── guidelines.pdf # Instructions used by human annotators for translating reviews ├── source_data/ │ ├── benchmark/ # Raw source reviews (TXT) for both fashion and electronics domains │ └── training/ # Raw source reviews (TXT) for both fashion and electronics domains └── translated_data/ ├── benchmark/ # Translated reviews (CSV) across 8 target languages └── training/ # Translated reviews (CSV) across 8 target languages ``` ## Data Format All CSV files follow the same column structure: | Column | Description | | :--- | :--- | | `src` | The English product review (Source Text). | | `tgt` | The corresponding human-annotated, style-aware target language translation. | | `src_lang` | Language code for Source (always `English`). | | `tgt_lang` | Language code for Target (e.g., `Hindi`, `Bengali`, `Tamil`, etc.). | | `domain` | Product category `E`(`electronics`) or `F`(`fashion`). | ## Quick Start You can easily load and explore the dataset using Python and Pandas: ``` import pandas as pd # Load the benchmark dataset for a specific language (e.g., Marathi) df_marathi = pd.read_csv("translated_data/benchmark/Marathi.csv") # View the first few style-aware translations print(df_marathi[['src', 'tgt', 'domain']].head()) ``` ## Citation If you use the **STAR-IL dataset**, please cite our paper: ```bibtex @inproceedings{shetye-etal-2026-star, title = {STAR-IL: A Dataset for Style-Aware Machine Translation of Product Reviews in Indian Languages}, author = {Shetye, Ketaki and Sharma, Dipti Misra and Krishnamurthy, Parameswari}, booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)}, month = may, year = {2026}, pages = {8780--8793}, address = {Palma, Mallorca, Spain}, publisher = {European Language Resources Association (ELRA)}, doi = {10.63317/4oq85vioi2tu}, abstract = {Product reviews on e-commerce platforms are a critical form of user-generated content influencing consumer decisions. However, these reviews are predominantly in English, creating accessibility barriers for non-fluent users. Existing translation models often fail to preserve domain-specific features and colloquial style, resulting in unnatural outputs. To address this, we introduce STAR-IL, a human-annotated multilingual parallel corpus for style-aware translation of product reviews. Experiments show that models fine-tuned on STAR-IL achieve an average improvement of 5.77 BLEU and 3.78 COMET points over baselines across all languages. The dataset is publicly available at https://github.com/ltrc/STAR-IL-Corpus.} }