Elamur Research⌕
☰
home / papers / 2026 / SHEINfer: Implicit Product Category Inference from Arabic E-commerce Reviews

SHEINfer: Implicit Product Category Inference from Arabic E-commerce Reviews

Paper · 2026 · Hend Al-Khalifa

Summary

The paper introduces SHEINfer, a task and dataset for guessing what kind of product an Arabic customer review is about when the review never names the product. Cues are indirect: comments on fabric, fit, charging speed or colour, often in Saudi/Gulf dialect. The author took reviews from the public Arabic Reviews of SHEIN collection on Hugging Face. Two native-speaker annotators independently labelled 801 of them into 11 categories using only the review text. The 515 reviews on which both annotators agreed became the benchmark.

Because the data is small, every system was scored with 5-fold stratified cross-validation. The systems were TF-IDF features with SVM or Logistic Regression, three Arabic BERT variants, and GPT-4o-mini prompted with zero or three examples. The dialect-oriented MARBERT reached the best accuracy, while the simple TF-IDF plus Logistic Regression baseline handled rare classes best by macro F1. The general-purpose LLM performed very poorly. The work sets a first baseline for inferring products from reviews in dialectal Arabic.

Key points

  • Source: Arabic Reviews of SHEIN (Bin Safi, 2024), 2,415 reviews with star ratings, mostly in Saudi dialect. Of these, 801 were double-annotated and 515 (64.3%) have full agreement, with Cohen's κ = 0.60.
  • There are 11 categories based on SHEIN's catalogue, including an explicit "Unclear" class. The distribution is skewed: Makeup/Beauty makes up 25.8%, while Children's Clothing has only 5 items (1.0%).
  • MARBERT achieved the top accuracy (0.586 ± 0.026). TF-IDF + Logistic Regression achieved the top macro F1 (0.417 ± 0.056). AraBERT and CAMeLBERT were also tested.
  • GPT-4o-mini reached only 0.064 accuracy zero-shot and 0.186 with 3 examples.
  • Stated limitations:
    • only one LLM was tested;
    • no chain-of-thought or self-consistency prompting was tried;
    • the data covers a single platform and dialect.

Interesting ideas

  • I1. The label set includes a category for reviews where even humans cannot determine the product, so the model can abstain instead of being forced to guess.
  • I2. Only items where both independent annotators agree become gold, and the disagreements are analysed as signs of real ambiguity, e.g. "comfortable" could fit clothing or shoes.
  • I3. The paper sorts implicit cues into four types: quality descriptors, fit/size, function, and aesthetics, each pointing to a product type with different strength.
  • I4. Results are reported as both accuracy and macro F1 with fold-level standard deviation, and the two metrics produce different winners on imbalanced data.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (answer): Readers could have an explicit "the text does not say" outcome, trained and scored as a legitimate answer rather than treated as a failure.
  • I2 (learn): Silver data from witness parsers could follow the same rule, with the non-agreed sentences kept as a separate diagnostic set of genuinely ambiguous constructions.
  • I3 (world): An unnamed entity's type could be inferred from attributes predicated of it. For example, "charging is fast" implies an object with a battery, using the WordNet/Wikidata ontology, and the chain of reasoning stays explainable.
  • I4 (learn): Evaluations on skewed label sets, such as rare dependency relations or rare frames, should always report macro scores and variance so that majority-class gains do not hide regressions.

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

 Institutions

🎓King Saud UniversityRiyadh · Saudi Arabia37 works · 527 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Hend Al-Khalifa. 2026. SHEINfer: Implicit Product Category Inference from Arabic E-commerce Reviews. The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4t8cfnk4ki9a

@inproceedings{alkhalifa2026sheinfer,
  title = {{SHEINfer: Implicit Product Category Inference from Arabic E-commerce Reviews}},
  author = {Hend Al-Khalifa},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/4t8cfnk4ki9a},
  url = {https://doi.org/10.63317/4t8cfnk4ki9a}
}

References

19 references, 4 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., Elmadany, A., & Nagoudi, E. M. B. (2021). ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 7088-7105. ↗
  2. Al-Smadi, M., Qawasmeh, O., Al-Ayyoub, M., Jararweh, Y., & Gupta, B. (2018). Deep Recurrent Neural Network vs. Support Vector Machine for Aspect-Based Sentiment Analysis of Arabic Hotels' Reviews. Journal of Computational Science, 27, 386-393. ↗
  3. Antoun, W., Baly, F., & Hajj, H. (2020). AraBERT: Transformer-based Model for Arabic Language Understanding. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT), pages 9-15. ↗
  4. Bin Safi, R. (2024). Arabic Reviews of SHEIN. Hugging Face Datasets. https://huggingface.co/datasets/Ruqiya/Arabic _Reviews_of_SHEIN Blume, A., Zalmout, N., Ji, H., & Li, X. (2023). Generative Models for Product Attribute Extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 575-585. ↗
  5. Brown, T., Mann, B., Ryder, N., Subbiah, M., et al. (2020). Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems (NeurIPS), 33, pages 1877-1901. ↗
  6. Çiftlikçi, M. S., Çakmak, Y., Kalaycı, T. A., Abut, F., Akay, M. F., & Kızıldağ, M. (2025). A New Large Language Model for Attribute Extraction in E-Commerce Product Categorization. Electronics, 14(10), 1930. ↗
  7. Elnagar, A., Khalifa, Y. S., & Einea, A. (2018). Hotel Arabic-Reviews Dataset Construction for Sentiment Analysis Applications. In Intelligent Natural Language Processing: Trends and Applications, pages 35-52. Springer. ↗
  8. Fang, C., Li, X., Fan, Z., Xu, J., Nag, K., Korpeoglu, E., Kumar, S., & Achan, K. (2024). LLM-Ensemble: Optimal Large Language Model Ensemble Method for E-commerce Product Attribute Value Extraction. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. find
  9. Habbat, N., Anoun, H., & Hassouni, L. (2023). Enhancing Arabic Sentiment Analysis in E-Commerce Reviews on Social Media Through a Stacked Ensemble Deep Learning Approach. Mathematical Modelling of Engineering Problems, 10(3), 867-876. find
  10. Inoue, G., Alhafni, B., Baimber, N., Bouamor, H., & Habash, N. (2021). The Interplay of Variant, Size, and Task Type in Arabic Pre-trained Language Models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop (WANLP), pages 92-104.
  11. Landis, J. R., & Koch, G. G. (1977). The Measurement of Observer Agreement for Categorical Data. Biometrics, 33(1), 159-174. ↗
  12. Lin, R., He, X., Feng, J., Zalmout, N., Liang, Y., Xiong, L., & Dong, X. L. (2021). PAM: Understanding Product Images in Cross Product Category Attribute Extraction. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 3192-3200. find
  13. Wang, Q., Yang, L., Kanagal, B., Sanghai, S., Sivakumar, D., Shu, B., Yu, Z., & Elsas, J. (2020). Learning to Extract Attribute Value from Product via Question Answering: A Multi-task Approach. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 47-55. find
  14. Yang, L., Wang, Q., Yu, Z., Kulkarni, A., Sanghai, S., Shu, B., Elsas, J., & Kanagal, B. (2022). MAVE: A Product Dataset for Multi-source Attribute Value Extraction. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, pages 1256-1265. ↗
  15. Zahera, H. M., & Sherif, M. A. (2020). ProBERT: Product Data Classification with Fine-tuned BERT Model. In Proceedings of the MIDAS Workshop at ESWC 2020. find
  16. Zheng, G., Mukherjee, S., Dong, X. L., & Li, F. (2018). OpenTag: Open Attribute Value Extraction from Product Profiles. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1049-1058. find
  17. Zhou, J., Liu, B., Acharya, J., Hong, Y., Lee, K.-C., & Wen, M. (2024). Enhanced E-Commerce Attribute Extraction: Innovating with Decorative Relation Correction and LLAMA 2.0-Based Annotation. In Proceedings of The Web Conference 2024. ↗
  18. Zou, H. P., Samuel, V., Zhou, Y., Zhang, W., Fang, L., Song, Z., Yu, P. S., & Caragea, C. (2024a). ImplicitAVE: An Open-Source Dataset and Multimodal LLMs Benchmark for Implicit Attribute Value Extraction. In Findings of the Association for Computational Linguistics: ACL 2024, pages 338-354. find
  19. Zou, H. P., Yu, G. H., Fan, Z., Bu, D., Liu, H., Dai, P., Jia, D., & Caragea, C. (2024b). EIVEN: Efficient Implicit Attribute Value Extraction using Multimodal LLM. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 6: Industry Track), pages 453-463. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/khalifa-2026-sheinfer-implicit-product-category-inference/