metadata

license: mit
task_categories:
  - text-classification
  - sentence-similarity
  - translation
language:
  - en
  - sr
tags:
  - cross-lingual
  - semantic-text-similarity
  - sts-benchmark
  - parallel-corpus
pretty_name: English-Serbian Semantic Text Similarity (STS) Benchmark
size_categories:
  - n<1K

Dataset Card for English-Serbian Semantic Text Similarity Benchmark

Dataset Description

Repository: [Provide a link to your dataset repository]

Dataset Summary

This dataset is a parallel English-Serbian Semantic Text Similarity (STS) benchmark. It was created to evaluate multilingual English-Serbian language models, with a focus on SBERT (Sentence-BERT) knowledge distillation. The dataset consists of sentence pairs in English and Serbian, along with their semantic similarity scores. The dataset uses the test split from the original STS benchmark. Only the second sentence of each pair was translated to Serbian. The translation was done automatically using the gpt-3.5-turbo-0125 model. The prompt used for the translation can be found in the repository.

Supported Tasks and Leaderboards

Semantic Similarity: The primary task for this dataset is to measure the semantic similarity between English-Serbian sentence pairs.
Cross-lingual Understanding: This dataset can be used to evaluate models' performance in understanding semantic relationships across English and Serbian languages.
Machine Translation: While not the primary focus, this dataset can also be used to assess English to Serbian translation quality.

Languages

The dataset contains text in two languages:

English (en)
Serbian (sr)

Dataset Structure

Data Instances

Each instance in the dataset consists of:'

Datast split
Genre (the domain the sentence belongs to)
Dataset (original dataset a setence is obtained from)
Year (which itteration of the STS dataset a sentence belogs to)
Sid (a unique sentence id for each sentence pair)
Sentence 1 (an original English sentence from the STS benchmark datast)
Sentence 2 in English (an original English sentence from the STS benchmark datast)
Setnece 2 in Serbian (an automatically translated Sentence 2 from the STS benchmark dataset)
A similarity score (kept from the original STS benchmark dataset)

Data Fields

en_sentence: The original English sentence
sr_sentence: The corresponding Serbian translation
similarity_score: A float value representing the semantic similarity between the two sentences

Data Splits

The dataset uses the test split from the original STS benchmark. Only the second sentence of each pair was translated to Serbian.

Dataset Creation

Curation Rationale

This dataset was created to address the need for multilingual semantic similarity benchmarks, specifically for the English-Serbian language pair. It aims to facilitate research in cross-lingual understanding and evaluation of multilingual language models.

Source Data

The source data used for the translation is the test split of the STS benchmark dataset from: https://huggingface.co/datasets/mteb/stsbenchmark-sts

Initial Data Collection and Normalization

The source data comes from the test split of the original STS benchmark dataset.

Considerations for Using the Data

Discussion of Biases

As the Serbian translations were produced automatically, there may be biases or errors introduced by the machine translation process. Users should be aware of potential gender, cultural, or linguistic biases that may be present in the original English sentences or introduced during translation.

Other Known Limitations

The dataset is limited to the test split of the original STS benchmark, which may affect its size and diversity.
The dataset may not cover all domains or types of semantic relationships.
Automatic translation may not capture all nuances of the original English sentences. A random split of 100 translated sentences was validated. However, translation does not involve post-correction at this stage. Thus, the translated Serbian sentences might contain inconsistencies in relation to the rules and grammar of the standard Serbian language.

Licensing Information

This dataset is published under the MIT license.

Citation Information

If you use this dataset, please link the smartcat huggingface dataset repository.

Loading the Dataset

To load this dataset using the Hugging Face datasets library, you can use the following code:

from datasets import load_dataset

# Replace 'username/dataset_name' with the actual path to your dataset on Hugging Face
dataset = load_dataset("smartcat/STS_parallel_en_sr")

# Access the data
for example in dataset['train']:
    print(f"Sentence 1: {example[sentence1]}")
    print(f"Sentence 2 in English: {example['sentence2_eng']}")
    print(f"Sentence 2 in Serbian: {example['sentence2_sr']}")
    print(f"Similarity Score: {example['score']}")
    print()