A comprehensive machine learning pipeline for detecting and classifying biotic interactions in scientific text using multiple state-of-the-art NLP models.

This project implements various classification approaches for identifying biotic interactions (predation, parasitism, pollination, etc.) from scientific literature passages. The system supports multiple model architectures and includes tools for training, evaluation, and inference.

• Multiple Model Architectures:

  • Transformer models (DistilBERT, BioBERT, RoBERTa, PubMedBERT)
  • Support Vector Machines (SVM)
  • Random Forest
  • Logistic Regression
  • BERT variants
  • LLaMA models
  • LUKE (entity-aware transformers)

• Advanced Training Features:

  • 5-fold stratified cross-validation
  • Active learning for efficient annotation
  • Precision-recall threshold optimization
  • High precision mode for reduced false positives
  • External validation on curated test sets

• Production-Ready APIs:

  • RESTful API endpoints for each model
  • Batch prediction support
  • Real-time inference

classifier/
├── src/                          # Source code
│   ├── models/                   # Model training scripts
│   │   ├── transformer_classifier.py    # Main transformer training pipeline
│   │   ├── svm_classifier.py            # SVM implementation
│   │   ├── random_forest_classifier.py  # Random Forest classifier
│   │   ├── bert_classifier.py           # BERT fine-tuning
│   │   ├── llama_classifier.py          # LLaMA model training
│   │   └── luke_classifier.py           # LUKE entity-aware model
│   ├── api/                      # API endpoints
│   │   ├── biomedbert_api.py
│   │   ├── svm_api.py
│   │   ├── lr_api.py
│   │   └── test_api.py
│   ├── data/                     # Data processing
│   │   ├── preprocessing.py
│   │   ├── data_building.py
│   │   ├── retrieve_positives.py
│   │   ├── retrieve_negatives.py
│   │   └── retrieve_random.py
│   └── utils/                    # Utilities
│       ├── prediction.py
│       ├── evaluation.py
│       └── classification_details.py
├── data/
│   ├── training/                 # Training datasets
│   ├── evaluation/               # Test/validation sets
│   ├── processed/                # Processed data
│   └── taxonomies/               # Species taxonomy data
├── models/                       # Trained models (git-ignored)
├── results/                      # Outputs
│   ├── cv_results/              # Cross-validation metrics
│   ├── predictions/             # Model predictions
│   └── figures/                 # Visualizations
├── notebooks/                    # Jupyter notebooks
└── scripts/                      # Utility scripts

git clone <your-repo-url>
cd classifier

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

pip install -r requirements.txt

# Download spaCy language models
python -m spacy download en_core_web_sm

Train all transformer models with cross-validation:

python src/models/transformer_classifier.py --model all --epochs 3 --cv_folds 5

Train a specific model:

python src/models/transformer_classifier.py --model biobert --epochs 5

Enable active learning:

python src/models/transformer_classifier.py --model BiomedBERT --active_learning --al_iterations 10

Optimize for high precision:

python src/models/transformer_classifier.py --model distilbert --target_precision 0.8

python src/models/svm_classifier.py

python src/utils/prediction.py --model biomedbert --input data/evaluation/eval_100.tsv

python src/api/biomedbert_api.py

Results on external test set (eval_100.tsv):

See results/cv_results/ for detailed metrics.

• training_data_cleaned.csv: Cleaned training set with labeled passages
• Format: passage,label where label is 0 (no interaction) or 1 (interaction)

• eval_100.tsv: Hand-curated test set (100 samples)
• BiotXBench annotations: Additional quality-checked annotations
• GloBI datasets: Global Biotic Interactions passages

The transformer classifier supports uncertainty-based active learning:

1. Start with small labeled dataset (default: 500 samples)
2. Train initial model
3. Select most uncertain samples from unlabeled pool
4. Iteratively expand training set
5. Track performance improvement over iterations

Example:

python src/models/transformer_classifier.py \
  --model biobert \
  --active_learning \
  --al_iterations 10

For high-precision applications (e.g., automated curation), the system can optimize decision thresholds:

python src/models/transformer_classifier.py \
  --model BiomedBERT \
  --target_precision 0.9

This finds the optimal probability threshold to achieve target precision while maximizing recall.

POST /predict
{
  "text": "The fox hunts mice in the meadow."
}

Response:
{
  "prediction": 1,
  "confidence": 0.94,
  "label": "interaction"
}

POST /predict_batch
{
  "texts": ["sentence 1", "sentence 2", ...]
}

esteban.gaillac@hesge.ch