AOSED — Arabic Opinion Summary Evaluation Dataset

AOSED is the first evaluation benchmark for Arabic opinion summarization. It pairs Arabic customer reviews of 100 restaurants with human-validated reference summaries, summaries generated by three large language models, and human annotations of summary quality.

The dataset accompanies the paper ArabicOpinion: A Five-Dimensional Multi-Granular Evaluation Framework for Arabic Opinion Summarization, published in ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP).

Contents

	Count
Restaurants	100
Customer reviews	1,944 (10–27 per restaurant, mean 19.44)
Human-validated reference summaries	100
LLM-generated summaries	300 (GPT-4o, Claude, JAIS)
Annotated summaries	300, each scored on relevance, faithfulness, and sentiment
Gold opinion units	686
Gold entity units	249
System units	3,312

Files

restaurants.csv — place_id, average_rating

reviews.csv — place_id, review_id, review_text One row per review. Text only; no reviewer names, usernames, or identifiers. Line breaks within a review are preserved as written.

gold_summaries.csv — place_id, golden_summary, sentiment_gold, n_reviews, n_opinions, n_entities, n_opinion_units_extracted, n_entity_units_extracted

n_opinions and n_entities are the counts recorded by the annotators; these are the denominators used to compute the published relevance scores. The _extracted columns are the counts recoverable from the unit text. They agree for 98 of 100 restaurants; in the two exceptions the difference is a single unit.

gold_units.csv — place_id, unit_id, unit_type, unit_text Opinion phrases and named entities extracted from each reference summary. unit_type is opinion or entity.

system_summaries.csv — place_id, model, summary, sentiment_system, n_repetitions, n_conflicts n_repetitions and n_conflicts are annotator counts of repeated and mutually contradictory opinions within a summary.

system_units.csv — place_id, model, unit_id, unit_type, unit_text

annotations.csv — place_id, model, relevance, faithfulness Final human scores in [0, 1]. Relevance is the proportion of reference opinions and entities covered by the system summary. Faithfulness is the proportion of system opinions and entities supported by the reference summary.

aosed_full.csv — a flat 300-row join of the above, for convenience.

statistics.csv — summary statistics for the whole dataset.

Tables are linked by place_id, and the system-side tables additionally by model.

How it was built

Reviews were collected from Google Maps for 100 restaurants. Reference summaries were produced with a hybrid approach: GPT-4o generated initial drafts, which three qualified native-Arabic annotators then corrected, rewrote, and validated over three rounds against criteria covering fluency, coherence, relevance, factuality, aspect coverage, sentiment consistency, specificity, entity coverage, redundancy, and the handling of conflicting opinions.

Three LLMs generated summaries for the same inputs at temperature 0.1: GPT-4o, Claude 3 Sonnet (claude-3-sonnet-20240229), and JAIS-30B (inceptionai/jais-family-30b-16k-chat).

Relevance and faithfulness were annotated by counting covered and supported units, under a structured consensus protocol: annotators worked independently, then resolved disagreements by discussion and majority vote. Sentiment was labelled with a binary scheme by three human annotators plus GPT-4o as a fourth voice, reaching Fleiss' Kappa 0.670 among the human annotators.

Full methodology is in the paper. The guidelines given to the annotators for constructing the reference summaries are included in this repository.

Licensing and intended use

Released under CC BY-NC-SA 4.0 for non-commercial research.

The reference summaries, annotations, extracted units, and LLM-generated summaries are the authors' contribution. The customer review texts are publicly posted user content collected from Google Maps for non-commercial research; the authors claim no ownership of them.

Citation

@article{aldashnan2026arabicopinion,
  title   = {ArabicOpinion: A Five-Dimensional Multi-Granular Evaluation Framework
             for Arabic Opinion Summarization},
  author  = {Aldashnan, Bayan and Alothaim, Abdulrahman and Alsanad, Ahmed},
  journal = {ACM Transactions on Asian and Low-Resource Language Information Processing},
  year    = {2026},
  publisher = {Association for Computing Machinery}
}