README.md

metadata

dataset_info:
  features:
    - name: sentence1
      dtype: string
    - name: sentence2
      dtype: string
    - name: score
      dtype: float64
    - name: langs
      dtype: string
  splits:
    - name: s1_ar_ar
      num_bytes: 2368220
      num_examples: 11512
    - name: s2_en_en
      num_bytes: 1615474
      num_examples: 11512
    - name: s3_multilingual_1
      num_bytes: 1917019
      num_examples: 5756
    - name: s4_multilingual_2
      num_bytes: 1917019
      num_examples: 5756
  download_size: 3993518
  dataset_size: 7817732
configs:
  - config_name: default
    data_files:
      - split: s1_ar_ar
        path: data/s1_ar_ar-*
      - split: s2_en_en
        path: data/s2_en_en-*
      - split: s3_multilingual_1
        path: data/s3_multilingual_1-*
      - split: s4_multilingual_2
        path: data/s4_multilingual_2-*
license: apache-2.0
task_categories:
  - sentence-similarity
language:
  - ar
  - en
size_categories:
  - 10K<n<100K

SILMA STS Arabic/English Dataset - v1.0

Overview

The SILMA STS Arabic/English Dataset - v1.0 is a dataset designed for training and evaluating sentence embeddings for Arabic and English tasks. It consists of five different splits that cover monolingual and multilingual sentence pairs, with human-annotated similarity scores. The dataset includes both Arabic-to-Arabic and English-to-English pairs, as well as cross-lingual Arabic-English pairs, making it a valuable resource for multilingual and cross-lingual semantic similarity tasks.

Dataset Structure

The dataset is divided into five splits, each containing sentence pairs and similarity scores.

Split 1: ar_ar

• Description: Contains Arabic-to-Arabic sentence pairs with similarity scores.
• Size: 11,512 examples
• JSON Sample:

  {
    "sentence1": "رجلين يلعبان الشطرنج",
    "sentence2": "ثلاثة رجال يلعبون الشطرنج",
    "score": 0.52,
    "langs": "ar-ar"
  }
  

Split 2: en_en

• Description: Contains English-to-English sentence pairs with similarity scores.
• Size: 11,512 examples
• JSON Sample:

  {
    "sentence1": "A plane is taking off.",
    "sentence2": "An air plane is taking off.",
    "score": 1.0
  }
  

Split 3: multilingual_1

• Description: Contains sentence pairs from both Arabic and English, with similarity scores. The sentences are aligned cross-lingually.
• Size: 5,756 examples
• JSON Sample:

  {
    "sentence1": "The man is playing the guitar. | الرجل يعزف على الغيتار",
    "sentence2": "The man is playing the piano. | الرجل يعزف على البيانو",
    "score": 0.32
  }
  

Split 4: multilingual_2

• Description: Similar to Split 3, but with reversed language pairs.
• Size: 5,756 examples
• JSON Sample:

  {
    "sentence1": "رجل يدخن | A man is smoking.",
    "sentence2": "رجل يتزلج | A man is skating.",
    "score": 0.1
  }
  

Column Descriptions

Each split in the dataset contains the following columns:

• sentence1: The first sentence in the pair. It can be in Arabic or English depending on the split.
• sentence2: The second sentence in the pair. It can also be in Arabic or English depending on the split.
• score: A floating-point number between 0 and 1 representing the semantic similarity between the two sentences, where 1 indicates maximum similarity.
• langs: Indicates the language pair of the sentences. The possible values are:
  • ar-ar (Arabic-Arabic)
  • en-en (English-English)
  • Multilingual-1 (Multilingual, English-Arabic)
  • Multilingual-2 (Multilingual, Arabic-English)

Use Cases

The SILMA STS Arabic/English Dataset - v1.0 can be used in various NLP tasks, including but not limited to:

1. Sentence Embedding Training: The dataset is well-suited for training models that generate sentence embeddings, enabling effective comparison of sentence-level semantics in both Arabic and English.
2. Multilingual and Cross-Lingual STS: This dataset can be used for evaluating the performance of multilingual and cross-lingual sentence transformers, as it includes both monolingual and multilingual sentence pairs.
3. Semantic Similarity Tasks: The dataset can be utilized in semantic similarity benchmarks, particularly for Arabic and English language pairs.
4. Cross-Lingual Transfer Learning: The multilingual sentence pairs provide a good opportunity for training models in cross-lingual transfer learning, where knowledge from one language can be transferred to another.

This dataset is a useful resource for researchers and developers working on NLP tasks that involve sentence semantics across different languages, especially for Arabic and English.