RuDReC

RuDReC provides a Russian-language corpus and neural models for extracting drug- and disease-related named entities and for detecting mentions of drug reactions and perceived effectiveness in user reviews.

Key Features:

• Corpus Composition: A raw corpus of 1.4 million health-related texts from internet platforms in Russian and a labeled subset of 500 consumer reviews annotated at sentence and expression levels, including drug classes, forms, indications, and reactions.
• NER and Sentence Classification: Baseline models for named entity recognition and multilabel sentence classification, including RuDR-BERT which achieves a macro F1 of 74.85% for NER and 68.82% for sentence classification, a 7.47% improvement over a standard Russian BERT.
• Domain-Specific Pretrained Models: Domain-specific BERT models pretrained on the RuDReC corpus to improve processing of Russian health-related text.

Scientific Applications:

• Information Extraction (IE): Extraction of drug- and disease-related entities and statements from user reviews to support comparison with traditional sources such as drug labels.
• Adverse Drug Reaction Detection: Identification and monitoring of patient-reported adverse effects mentioned in consumer reviews.
• Drug Effectiveness Analysis: Analysis of user-reported perceptions of pharmaceutical effectiveness described in reviews.

Methodology:

Training and evaluation of tailored BERT variants (including RuDR-BERT and domain-specific pretrained BERTs) on the 1.4 million-text raw corpus and the 500-review labeled subset for named entity recognition and multilabel sentence classification with reporting of macro F1 metrics.