transformers-sklearn

transformers-sklearn provides an interface to fine-tune and apply transformer models for medical natural language processing tasks including classification, named entity recognition, and regression.

Key Features:

• Primary methods: Implements the functions fit, score, and predict for training, evaluation, and inference with transformer models.
• Model configuration: Exposes customizable parameters such as model_name_or_path and model_type for selecting pretrained transformer checkpoints and architectures.
• Multilingual support: Supports multilingual NLP tasks across different language datasets.
• Input data handling: Automatically generates required input data formats from annotated corpora.
• Predefined frameworks and training methods: Provides predefined model frameworks and training methods for common NLP tasks.

Scientific Applications:

• TrialClassification: Multi-label classification of Chinese medical trial texts, achieving a macro F1 score of 0.8225.
• BC5CDR: English biomedical named entity recognition, achieving a macro F1 score of 0.8703.
• DiabetesNER: Chinese diabetes entity recognition, achieving a macro F1 score of 0.6908.
• BIOSSES: English biomedical sentence similarity estimation, achieving a Pearson correlation of 0.8260.

Methodology:

Implements three Python classes—BERTologyClassifier, BERTologyNERClassifier, and BERTologyRegressor—each providing methods to fine-tune transformer-based models (fit), evaluate performance (score), and predict labels on test data (predict).