Hierarchical Multi-Label Classifier of Patent Abstracts by International Patent Classification

The tool supports research into innovation, technology diffusion, and regional economic development by automatically assigning patents to their Section, Class, and Subclass categories, enabling large-scale analysis of patenting activity that would be impractical to classify manually.<br><br>The classifier enforces the taxonomic structure of the IPC hierarchy during prediction: Section-level predictions constrain Class-level predictions, which in turn constrain Subclass-level predictions, ensuring that outputs are always taxonomically valid.

Patents can be assigned multiple labels at each level to reflect real-world multi-category patents. The underlying model fine-tunes SciBERT, a transformer model pretrained on scientific text, to improve classification accuracy on technical patent language, with mixed-precision (FP16) training support for efficient use of GPU resources on HPC systems.<br><br>Key features include configurable soft or hard constraint masking between hierarchy levels, multi-level evaluation metrics, and support for both local training/inference and batch submission via SLURM on HPC clusters.</p><p dir="ltr">IPC-Classifier is built in Python 3.10+ using PyTorch 2.0+ and Hugging Face Transformers 4.53+, with code quality enforced via Black, isort, and Pylint pre-commit hooks.</p><p dir="ltr">The software and codebase was developed by Dr.