--- license: apache-2.0 language: - ar - en task_categories: - text-retrieval - sentence-similarity pretty_name: Arabic Dialect-to-MSA Embedding Test Data --- # Arabic Dialect-to-MSA Embedding Test Dataset ## Overview This dataset evaluates embedding models on the retrieval of Modern Standard Arabic (MSA) passages from queries written in MSA and five Arabic dialects. Each original MSA query has five dialect variants: Saudi, Syrian, Egyptian, Sudanese, and Tunisian. The dataset is designed to measure retrieval performance across these query languages and compare base and fine-tuned embedding models. ## Dataset Details | Property | Value | |---|---| | Number of rows | 1,238 | | Original query language | MSA | | Target passage language | MSA | | Dialect variants | Saudi, Syrian, Egyptian, Sudanese, Tunisian | | Positive passages per row | 1 | | Hard negative passages per row | 9 | ## Dataset Structure Each row contains: - `query`: Original MSA query. - `positive`: Relevant MSA passage. - `negative1` to `negative9`: Nine hard negative MSA passages. - `saudi_variant`: Saudi Arabic query. - `syrian_variant`: Syrian Arabic query. - `egyptian_variant`: Egyptian Arabic query. - `sudanese_variant`: Sudanese Arabic query. - `tunisian_variant`: Tunisian Arabic query. ## Evaluation The dataset can be used to evaluate semantic retrieval using metrics such as: - **MRR@5:** Ranking quality of the first relevant result in the top five. - **Recall@5:** Proportion of queries with a relevant result in the top five. Evaluate each query language separately to measure the performance gap between MSA and dialect queries. ## Data Generation The original queries are in MSA. Dialect variants are generated using: 1. `facebook/nllb-200-1.3B` for MSA-to-English translation. 2. `oddadmix/Jisr-MT-50M-AllDialects` for English-to-dialect translation. ## Related Resources - [Fine-tuned model](https://huggingface.co/OussamaBenSlama/e5-small-dialect2MSA) - [Training and validation dataset](https://huggingface.co/datasets/OussamaBenSlama/e5-dialect2MSA-train-embedding-data) - [Base model](https://huggingface.co/intfloat/multilingual-e5-small) ## Limitations The dialect variants are machine-translated. They may not represent all natural dialect expressions. The dataset covers five Arabic dialects and is intended for MSA passage retrieval. ## License Apache-2.0.