On Using Arabic Language Dialects in Recommendation Systems

Abdulla Alshabanah, Murali AnnavaramNAACL 2025 · DOI

Abstract: Arabic is a diglossic language where formal writing (Modern Standard Arabic) differs significantly from regional dialects used in daily life. This paper presents the first attempt to incorporate Arabic dialects as a signal in recommendation systems. By leveraging Arabic Dialect Identification (ADI) and the Arabic Level of Dialectness (ALDi) from user reviews, the researchers achieved significant performance boosts—up to a 27% improvement in Hit Rate—demonstrating that linguistic variation is a powerful proxy for user preference and identity.

Introduction

Arabic is spoken by over 420 million people, but it is not a linguistic monolith. It exists in a state of diglossia: Modern Standard Arabic (MSA) is used for formal writing, while numerous regional dialects are used in daily life and social media. These dialects differ so much in grammar and vocabulary that they can be mutually unintelligible.

In recommendation systems, user reviews are often analyzed to understand preferences. However, most NLP tools for Arabic focus on MSA. Dialectal Arabic is often treated as "low-resource" noise due to its lack of standardized spelling and frequent code-switching. The authors argue that these linguistic choices are "identity markers" that reflect a user’s cultural background and specific tastes. Ignoring these dialects means leaving valuable personalized data on the table.

Methodology

The authors built their approach on a standard Two-Tower Neural Network (TTNN) architecture, exploring two main ways to inject dialect information:

Explicit Signal

The researchers extracted dialect labels (e.g., Egyptian, Levantine) and dialect intensity scores (ALDi) from user reviews using BERT-based models. This information was fed directly into the user tower as a feature. They tested variations including single-label and multi-label approaches, finding that multi-label representations better captured the nuances of code-switching.

Implicit Signal

Instead of feeding a label, they used the ALDi score (how "dialectal" a review is) to influence how the model perceives a user's history. They sampled items from a user's past interactions with a probability proportional to the dialectness of the associated review. The intuition is that reviews written in heavy dialect are more "honest" or representative of the user's true identity. They tested Implicit_uniform (a simple mean of sampled item representations) and Implicit_weighted (using an attention mechanism).

Key Results and Insights

The researchers compared their models against a "No-Dialect" baseline using Hit Rate (HR@10) and Normalized Discounted Cumulative Gain (NDCG@10) across two datasets: BRAD (Book Reviews in Arabic Dataset) and LABR (Large-scale Arabic Book Reviews Dataset).

The Implicit_uniform model was the clear winner on the BRAD dataset, improving HR@10 by 27% and NDCG@10 by 30% over the baseline. On the LABR dataset, the Implicit_weighted model showed a 19% improvement in HR@10.

Comparison of Hit Rate (HR@10) for implicit models using random vs. ALDi-biased sampling.

A key insight is that linguistic style serves as a proxy for cultural identity. A user writing in a specific Saudi dialect likely has different preferences than one writing in a Moroccan dialect. Interestingly, passing raw review text through standard sentence transformers actually degraded performance, suggesting that the "noise" of inconsistent dialectal spelling confuses standard models, whereas extracting the level of dialectness provides a cleaner, more useful signal.

Comparison of NDCG@10 for implicit models using random vs. ALDi-biased sampling.

Conclusion

This work demonstrates that Arabic dialects are not just a challenge for NLP but a powerful signal for personalization. By treating dialectal variation as an identity marker rather than noise, recommendation systems can achieve significant performance gains. Future work will explore jointly fine-tuning dialect identification models with recommendation systems to further align linguistic signals with user preferences.

Summary updated

BibTeX
@inproceedings{alshabanah2025using,
  author = {Abdulla Alshabanah and Murali Annavaram},
  title = {{On Using Arabic Language Dialects in Recommendation Systems}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-naacl.115},
}
Figure