The hospitality industry in the Arab world increasingly relies on customer feedback to shape services, driving the need for advanced Arabic sentiment analysis tools. To address this challenge, the Sentiment Analysis on Arabic Dialects in the Hospitality Domain shared task focuses on Sentiment Detection in Arabic Dialects. This task leverages a multi-dialect, manually curated dataset derived from hotel reviews originally written in Modern Standard Arabic (MSA) and translated into Saudi and Moroccan (Darija) dialects. The dataset consists of 538 sentiment-balanced reviews spanning positive, neutral, and negative categories. Translations were validated by native speakers to ensure dialectal accuracy and sentiment preservation. This resource supports the development of dialect-aware NLP systems for real-world applications in customer experience analysis. More than 40 teams have registered for the shared task, with 12 submitting systems during the evaluation phase. The top-performing system achieved an F1 score of 0.81, demonstrating the feasibility and ongoing challenges of sentiment analysis across Arabic dialects.

This paper details the iWAN-NLP system developed for participation in the AHaSIS 2025 shared task, “Sentiment Analysis on Arabic Dialects in the Hospitality Domain: A Multi-Dialect Benchmark.” Our approach leverages a multi-model ensemble strategy, combining the strengths of MARBERTv2, Saudibert, and DarijaBERT. These pre-trained Arabic language models were fine-tuned for sentiment classification using a 5-fold stratified cross-validation methodology. The final predictions on the test set were derived by averaging the logits produced by each model across all folds and then averaging these combined logits across the three models. This system achieved a macro F1-score of 81.0% on the official evaluation dataset and a cross-validated macro F1-score of 0.8513 (accuracy 0.8628) on the training set. Our findings highlight the effectiveness of ensembling regionally adapted models and robust cross-validation for Arabic sentiment analysis in the hospitality domain, ultimately securing first place in the AHaSIS 2025 shared task.

Arabic exhibits a rich and intricate linguistic landscape, with Modern Standard Arabic (MSA) serving as the formal written and spoken medium, alongside a wide variety of regional dialects used in everyday communication. These dialects vary considerably in syntax, vocabulary, phonology, and meaning, presenting significant challenges for natural language processing (NLP). The complexity is particularly pronounced in sentiment analysis, where emotional expressions and idiomatic phrases differ markedly across regions, hindering consistent and accurate sentiment detection. This paper describes our submission to the Ahasis Shared Task: A Benchmark for Arabic Sentiment Analysis in the hospitality domain. This shared task focuses on advancing sentiment analysis techniques for Arabic dialects in the hotel domain. Our proposed approach achieved an F1 score of 0.88 % on the internal test set (split from the original training data), and 79.16% on the official hidden test set of the shared task. This performance secured our team second place in the Ahasis Shared Task.

This work addresses the challenge of Arabic sentiment analysis in the hospitality domain in all dialects by using data augmentation techniques. We created a pipeline with three simple techniques: context-based paraphrasing, pattern-based sentence generation, and domain-specific word replacement. Our method preserves the original dialect features, meanings, and key classification details while adding diversity to the training data. It also includes automatic fallback between methods to handle challenges effectively. We used the Fanar API for dialectal data augmentation in the hospitality domain. The AraBERT-Large-v02 model was fine-tuned on original and augmented data, showing improved performance. This study helps solve the problem of limited dialect data in Arabic NLP and offers an effective framework that is useful for other Arabic text analysis tasks.

This work was conducted as part of the Ahasis@RANLP–2025 shared task, which focuses on sentiment detection in Arabic dialects within the hotel review domain. The primary objective is to advance sentiment analysis methodologies tailored to dialectal Arabic. Our work combines data augmentation with a hybrid model that integrates AraBERT and our created sentiment lexicon. Notably, our hybrid model significantly improved performance, reaching an F1-score of 0.74, compared to 0.56 when using only AraBERT. These results highlight the effectiveness of lexicon integration and augmentation strategies in enhancing both the accuracy and robustness of sentiment classification in dialectal Arabic.

In this paper, we describe our contribution in Ahasis shared task: Sentiment analysis on Arabic Dialects in the Hospitality Domain. Through the presented framework, we explored using two learning strategies tailored to a Large Language Model (LLM) and Transformer-based model variants. While few-shot prompting was used with GPT-4o, fine-tuning was adopted once to refine the essential MARBERT model on the Ahasis dataset and then to utilize a MARBERT variant model, SODA-BERT, that was pretrained on an Omani sentiment dataset and later evaluated with the shared task data.

This paper presents our approach to Arabic sentiment analysis with a specific focus on dialect-awareness for Saudi and Moroccan (Darija) dialectal variants. We develop a system that achieves a macro F1 score of 77% on the test set, demonstrating effective generalization across these dialect variations. Our approach leverages a pre-trained Arabic language model (Qarib) with custom dialect-specific embeddings and preprocessing techniques tailored to each dialect. The results show a significant improvement over baseline models that do not incorporate dialect information, with an absolute gain of 5% in F1 score over the equivalent non-dialect-aware model. Our analysis further reveals distinct sentiment expression patterns between Saudi and Darija dialects, highlighting the importance of dialect-aware approaches for Arabic sentiment analysis.

Sentiment analysis of Arabic dialects presents significant challenges due to linguistic diversity and the scarcity of annotated data. This paper describes our approach to the AHaSIS shared task, which focuses on sentiment analysis on Arabic dialects in the hospitality domain. The dataset comprises hotel reviews written in Moroccan and Saudi dialects, and the objective is to classify the reviewers’ sentiment as positive, negative, or neutral. We employed the SetFit (Sentence Transformer Fine-tuning) framework, a data-efficient few-shot learning technique. On the official evaluation set, our system achieved an F1 of 73%, ranking 12th among 26 participants. This work highlights the potential of few-shot learning to address data scarcity in processing nuanced dialectal Arabic text within specialized domains like hotel reviews.

Sentiment Analysis is a crucial task in Natural Language Processing (NLP) focused on identifying and categorizing emotional tones or opinions within text. For Arabic customer reviews, sentiment analysis is particularly challenging. The language’s rich diversity, with numerous regional dialects differing significantly from Modern Standard Arabic (MSA) and each other in lexicon, syntax, and sentiment expression, complicates consistent performance across dialects. In this paper, we present our approach, submitted to the AHASIS Shared Task 2025, focusing on sentiment analysis for Arabic dialects in the hotel domain. Our method leverages the capabilities of GPT-4o through adaptive few-shot prompting technique, where similar contextual examples are dynamically selected for each review using a k-Nearest Neighbors (kNN) search over train embeddings from a fine-tuned encoder model. This approach tailors the prompt to each specific instance, enhancing classification performance over minority class. Our submission achieved an F1-score of 76.0% on the official test set, showing stronger performance for the Saudi dialect compared to Darija.

This paper describes the AraNLP system developed for the “Ahasis” shared task on sentiment detection in Arabic dialects for hotel reviews. The task involved classifying the overall sentiment of hotel reviews (Positive, Negative, or Neutral) written in Arabic dialects, specifically Saudi and Darija. Our proposed model, AraNLP, is a hybrid deep learning classifier that leverages the strengths of a transformer-based Arabic model (AraELECTRA)augmented with classical bag-of-words style features (TF-IDF). Our system achieved an F1-score of 76%, securing the 5th rank in the shared task, significantly outperforming the baseline system’s F1-score of 56%.

This paper presents a study on sentiment anal- ysis of Dialectal Arabic (DA), with a particu- lar focus on Saudi and Moroccan (Darija) di- alects within the hospitality domain. We in- troduce a novel dataset comprising 698 Saudi Arabian proverbs annotated with sentiment polarity labels—Positive, Negative, and Neu- tral—collected from five major Saudi dialect regions: Najdi, Hijazi, Shamali, Janoubi, and Sharqawi. In addition to this, we used customer reviews for fine-tuning the CAMeLBERT-DA- SA model, which achieved a 75% F1 score in sentiment classification. To further evaluate the robustness of Arabic-centric models, we assessed the performance of three open-source large language models—Allam, ACeGPT, and Jais—in a zero-shot setting using the Ahasis shared task test set. Our results highlight the effectiveness of domain-specific fine-tuning in improving sentiment analysis performance and demonstrate the potential of Arabic-centric LLMs in zero-shot scenarios. This work con- tributes new linguistic resources and empirical insights to support ongoing research in senti- ment analysis for Arabic dialect

This paper presents a sentiment analysis model tailored for Arabic dialects in the hospitality domain, developed for the Ahasis Shared Task. Leveraging the Gemini Pro 1.5 language model, we address the challenges posed by the diversity of Arabic dialects, specifically Saudi and Moroccan Darija. Our method used the official Ahasis dataset of 3,000 hotel reviews. Through iterative benchmarking, dialect labeling, sarcasm detection, and fine-tuning, we adapted Gemini Pro 1.5 for the task. The final model achieved an F1-score of 0.7361 and ranked 10th on the competition leaderboard. This work shows that prompt engineering and domain adaptation of LLMs can mitigate challenges of dialectal variation, sarcasm, and resource scarcity in Arabic sentiment classification. Our contribution lies in the integration of dialect-specific prompt tuning with real-time batch inference, avoiding retraining. This approach, validated across 3,000 competition samples and 700 internal benchmarks, establishes a novel template for Arabic-domain sentiment pipelines.

This paper presents our contribution to the AHASIS Shared Task at RANLP 2025, which focuses on sentiment analysis for Arabic dialects. While sentiment analysis has seen considerable progress in Modern Standard Arabic (MSA), the diversity and complexity of Arabic dialects pose unique challenges that remain underexplored. We address this by fine-tuning six pre-trained language models, including AraBERT, MARBERTv2, QARiB, and DarijaBERT, on a sentiment-labeled dataset comprising hotel reviews written in Saudi and Moroccan (Darija) dialects. Our experiments evaluate the models’ performance on both combined and individual dialect datasets. MARBERTv2 achieved the highest performance with an F1-score of 79% on the test set, securing third place among 14 participants. We further analyze the effectiveness of each model across dialects, demonstrating the importance of dialect-aware pretraining for Arabic sentiment analysis. Our findings highlight the value of leveraging large pre-trained models tailored to dialectal Arabic for improved sentiment classification.