Elamur Research⌕
☰
home / papers / 2024 / GLARE: Google Apps Arabic Reviews Dataset

GLARE: Google Apps Arabic Reviews Dataset

Summary

GLARE targets the lack of large, varied Arabic text resources. Many existing Arabic corpora come from Twitter, and dialectal Arabic is especially poorly served. The authors crawled the Saudi Google PlayStore with the google-play-scraper library (Node.js for app discovery, the Python port for reviews). They took the top 200 free apps in each of 59 main and sub categories. After deduplication, this gave 9,980 unique apps and about 76M reviews, around 69M of them in Arabic once filtering was applied. App metadata and review metadata are kept alongside the text, including ratings, thumbs-up counts, developer replies and app version.

The paper then offers a descriptive analysis of the data and some engineered features, such as a term-frequency vocabulary, review lengths, review counts per app and category combinations. Both raw and engineered versions are released on GitHub and Hugging Face. The authors claim it is the largest Arabic review collection so far. They see it as a source for sentiment analysis, ABSA, spam detection and software-engineering research. App reviews may run to 4,000 characters, which permits richer text than tweets.

Key points

  • Scale: 76,512,077 reviews in total (17 GB raw), about 69M Arabic after cleaning. They cover 9,980 apps, reduced from 11,736 by removing duplicates, across 59 categories and 213 distinct category combinations.
  • Ratings are heavily skewed toward the top. Over 80% of reviews (61.2M) give 5 stars, and 40.5% of reviews (about 30.9M) are a single word.
  • The CountVectorizer vocabulary holds 8.7M unique words, of which about 6.96M occur only once. Roughly 17% of the vocabulary is noise (non-Arabic-alphanumeric characters), and 15% of entries contain digits. The most frequent words are "jidan" (very, 12M), "mumtaz" (excellent, 5.6M) and "jameel" (beautiful, 5M).
  • Developers replied to about 3.8M reviews, and roughly 48% of apps engaged with reviewers. Azar leads with over 203K replies.
  • PUBG has the most reviews (3.2M), followed by YouTube (2.7M) and Facebook (2.5M), while 216 apps have only one review. Planned next steps are a domain-specific Arabic LM and benchmarks for ATE, ACD and sentiment.

Interesting ideas

  • I1. A star rating that users must give anyway can serve as free, distant supervision for opinion labels, which avoids manual annotation at scale. — For Mova (learn): naturally attached signals, such as answer keys or source metadata, could become weak labels for silver data. A negative control would check that the label is not simply recoverable from length or other surface cues.
  • I2. Profiling the vocabulary before cleaning reveals concrete filtering rules. The share of words seen once, the share of words containing digits or symbols, and extreme word lengths are all measured; every abnormally long "word" turned out to occur just once. — For Mova (tokens): a frequency-by-length audit of tokenizer output could flag garbage tokens, such as glued strings or emoji runs, before they reach the tagger and lemmatizer.
  • I3. The analysis makes the skew explicit: a large majority of five-star reviews and two in five reviews being a single word. This warns that naive training or scoring on such data rewards trivial behaviour. — For Mova (learn): evaluation sets could be stratified by sentence length and label so that a parser or reader cannot look good by handling only short, easy inputs.
  • I4. When duplicate app records disagreed on rating, the authors first dropped zero ratings, treating them as missing, then averaged the rest and kept one record. — For Mova (world): when entity records from Wikidata or text mentions are merged, sentinel values should be treated as missing before conflicting attributes are reconciled, rather than averaged in blindly.
  • I5. The dataset pairs user reviews with developer replies, and the authors analyse how this engagement relates to ratings. — For Mova (frames): complaint–response pairs like these are natural material for testing speech-act labels such as complaint, apology or promise, and for checking how a reply's acts relate to the review it answers.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2412.15259 · arXiv 2412.15259

@misc{alkhalifa2024glare,
  title = {{GLARE: Google Apps Arabic Reviews Dataset}},
  author = {Al-Khalifa;, Fatima AlGhamdi; Reem Mohammed; Hend and Alowisheq, Areeb},
  year = {2024},
  doi = {10.48550/arxiv.2412.15259},
  eprint = {2412.15259},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2412.15259}
}

References

17 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abdul-Mageed, M., Alhuzali, H., and Elaraby, M. (2018). You tweet what you speak: A city-level dataset of arabic dialects. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). find
  2. Ahmed, A., Ali, N., Alzubaidi, M., Zaghouani, W., Abd-alrazaq, A. A., and Househ, M. (2022). Freely available arabic corpora: A scoping review. Computer Methods and Programs in Biomedicine, 2:100049. find
  3. Al-Ayyoub, M., Khamaiseh, A. A., Jararweh, Y., and Al-Kabi, M. N. (2019). A comprehensive survey of arabic sentiment analysis. Information processing & management, 56(2):320–342. find
  4. Al-Shamani, M., Al-Sarem, M., Saeed, F., and Almutairi, W. (2022). Designing an arabic google play store user review dataset for detecting app requirement issues. In Advances on Smart and Soft Computing, pages 133–143. Springer. find
  5. Al-Smadi, M., Qawasmeh, O., Talafha, B., and Quwaider, M. (2015). Human annotated arabic dataset of book reviews for aspect based sentiment analysis. In 2015 3rd International Conference on Future Internet of Things and Cloud, pages 726–730. IEEE. find
  6. Al-Subaihin, A. A., Sarro, F., Black, S., Capra, L., and Harman, M. (2019). App store effects on software engineering practices. IEEE Transactions on Software Engineering, 47(2):300–319. find
  7. Alharbi, B., Alamro, H., Alshehri, M., Khayyat, Z., Kalkatawi, M., Jaber, I. I., and Zhang, X. (2020). Asad: A twitter-based benchmark arabic sentiment analysis dataset. arXiv preprint arXiv:2011.00578. ↗ our card
  8. Ali, Z. S., Mansour, W., Elsayed, T., and Al-Ali, A. (2021). Arafacts: the first large arabic dataset of naturally occurring claims. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 231–236. find
  9. Alsarsour, I., Mohamed, E., Suwaileh, R., and Elsayed, T. (2018). Dart: A large dataset of dialectal arabic tweets. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). find
  10. Aly, M. and Atiya, A. (2013). Labr: A large scale arabic book reviews dataset. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 494–498. find
  11. Alyafeai, Z., Masoud, M., Ghaleb, M., and Al-shaibani, M. S. (2021). Masader: Metadata sourcing for arabic text and speech data resources. arXiv preprint arXiv:2110.06744. D abrowski, J., Letier, E., Perini, A., and Susi, A. (2022). Analysing app reviews for software engineering: a systematic literature review. Empirical Software Engineering, 27(2):1–63. ↗ our card
  12. Einea, O., Elnagar, A., and Al Debsi, R. (2019). Sanad: Single-label arabic news articles dataset for automatic text categorization. Data in brief, 25:104076. find
  13. Elnagar, A., Khalifa, Y. S., and Einea, A. (2018). Hotel arabic-reviews dataset construction for sentiment analysis applications. In Intelligent Natural Language Processing: Trends and Applications, pages 35–52. Springer. find
  14. Haouari, F., Hasanain, M., Suwaileh, R., and Elsayed, T. (2020). Arcov-19: The first arabic covid-19 twitter dataset with propagation networks. arXiv preprint arXiv:2004.05861. ↗
  15. Mulki, H. and Ghanem, B. (2021). Let-mi: An arabic levantine twitter dataset for misogynistic language. arXiv preprint arXiv:2103.10195. ↗
  16. Nassif, A. B., Elnagar, A., Shahin, I., and Henno, S. (2021). Deep learning for arabic subjective sentiment analysis: Challenges and research opportunities. Applied Soft Computing, 98:106836. find
  17. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alghamdi-2024-glare-google-apps-arabic-reviews/