mlsa-iai-msu-lab/ru_sci_bench_mteb

RuSciBench Dataset Collection

This repository contains the datasets for the RuSciBench benchmark, designed for evaluating semantic vector representations of scientific texts in Russian and English.

Dataset Description

RuSciBench is the first benchmark specifically targeting scientific documents in the Russian language, alongside their English counterparts (abstracts and titles). The data is sourced from eLibrary.ru, the largest Russian electronic library of scientific publications, integrated with the Russian Science Citation Index (RSCI).

The dataset comprises approximately 182,000 scientific paper abstracts and titles. All papers included in the benchmark have open licenses.

Tasks

The benchmark includes a variety of tasks grouped into Classification, Regression, and Retrieval categories, designed for both Russian and English texts based on paper abstracts.

Classification Tasks

1. Topic Classification (OECD): Classify papers based on the first two levels of the Organization for Economic Co-operation and Development (OECD) rubricator (29 classes).
   • RuSciBenchOecdRuClassification (subset oecd_ru)
   • RuSciBenchOecdEnClassification (subset oecd_en)
2. Topic Classification (GRNTI/SRSTI): Classify papers based on the first level of the State Rubricator of Scientific and Technical Information (GRNTI/SRSTI) (29 classes).
   • RuSciBenchGrntiRuClassification (subset grnti_ru)
   • RuSciBenchGrntiEnClassification (subset grnti_en)
3. Core RISC Affiliation: Binary classification task to determine if a paper belongs to the Core of the Russian Index of Science Citation (RISC).
   • RuSciBenchCoreRiscRuClassification (subset corerisc_ru)
   • RuSciBenchCoreRiscEnClassification (subset corerisc_en)
4. Publication Type Classification: Classify documents into types like 'article', 'conference proceedings', 'survey', etc. (7 classes, balanced subset used).
   • RuSciBenchPubTypesRuClassification (subset pub_type_ru)
   • RuSciBenchPubTypesEnClassification (subset pub_type_en)

Regression Tasks

1. Year of Publication Prediction: Predict the publication year of the paper.
   • RuSciBenchYearPublRuRegression (subset yearpubl_ru)
   • RuSciBenchYearPublEnRegression (subset yearpubl_en)
2. Citation Count Prediction: Predict the number of times a paper has been cited.
   • RuSciBenchCitedCountRuRegression (subset cited_count_ru)
   • RuSciBenchCitedCountEnRegression (subset cited_count_en)

Retrieval Tasks

1. Direct Citation Prediction: Given a query paper abstract, retrieve abstracts of papers it directly cites from the corpus. Uses a retrieval setup (all non-positive documents are negative). (Dataset Link)
   • RuSciBenchCiteRuRetrieval
   • RuSciBenchCiteEnRetrieval
2. Co-Citation Prediction: Given a query paper abstract, retrieve abstracts of papers that are co-cited with it (cited by at least 5 common papers). Uses a retrieval setup. (Dataset Link)
   • RuSciBenchCociteRuRetrieval
   • RuSciBenchCociteEnRetrieval
3. Translation Search: Given an abstract in one language (e.g., Russian), retrieve its corresponding translation (e.g., English abstract of the same paper) from the corpus of abstracts in the target language. (Dataset Link)
   • RuSciBenchTranslationSearchEnRetrieval (Query: En, Corpus: Ru)
   • RuSciBenchTranslationSearchRuRetrieval (Query: Ru, Corpus: En)

Usage

These datasets are designed to be used with the MTEB library. First, you need to install the MTEB fork containing the RuSciBench tasks:

pip install git+https://github.com/mlsa-iai-msu-lab/ru_sci_bench_mteb

Then you can evaluate sentence-transformer models easily:

from sentence_transformers import SentenceTransformer
from mteb import MTEB

# Example: Evaluate on Russian GRNTI classification
model_name = "mlsa-iai-msu-lab/sci-rus-tiny3.1" # Or any other sentence transformer
model = SentenceTransformer(model_name)

evaluation = MTEB(tasks=["RuSciBenchGrntiRuClassification"]) # Select tasks
results = evaluation.run(model, output_folder=f"results/{model_name.split('/')[-1]}")

print(results)

For more details on the benchmark, tasks, and baseline model evaluations, please refer to the associated paper and code repository.

• Code Repository: https://github.com/mlsa-iai-msu-lab/ru_sci_bench_mteb
• Paper: https://doi.org/10.1134/S1064562424602191

Citation

If you use RuSciBench in your research, please cite the following paper:

@article{Vatolin2024,
  author  = {Vatolin, A. and Gerasimenko, N. and Ianina, A. and Vorontsov, K.},
  title   = {RuSciBench: Open Benchmark for Russian and English Scientific Document Representations},
  journal = {Doklady Mathematics},
  year    = {2024},
  volume  = {110},
  number  = {1},
  pages   = {S251--S260},
  month   = dec,
  doi     = {10.1134/S1064562424602191},
  url     = {https://doi.org/10.1134/S1064562424602191},
  issn    = {1531-8362}
}