ChemTables

ChemTables provides a curated dataset for semantic classification of tables extracted from chemical patents to support development and evaluation of machine learning models for table content categorization.

Key Features:

• Dataset composition: 788 labeled tables extracted from chemical patents, each annotated with a specific content-type label.
• Annotated content types: Annotations cover spectroscopic and physical properties, pharmacological uses, and effects of chemicals.
• Structural heterogeneity: Includes patent table layouts with images of Markush structures and merged cells that increase formatting variability.
• Task formulation: Defines a text-mining task for automated categorization of chemical patent tables by content type.
• Benchmark models: Baseline evaluations applied the neural architectures TabNet, ResNet, and Table-BERT.
• Performance: Table-BERT achieved a micro F1 score of 88.66 on the table classification task.

Scientific Applications:

• Information retrieval: Semantic classification of patent tables to improve retrieval of compound and reaction information from chemical patents.
• Data structuring for research: Structuring spectroscopic, physical, and pharmacological data within patent tables to support chemistry and pharmaceutical research.
• Model development: Development and evaluation of machine learning and NLP models for table content classification in scientific documents.

Methodology:

788 tables were extracted from chemical patents and annotated with content-type labels; benchmark experiments applied TabNet, ResNet, and Table-BERT, with Table-BERT evaluated using micro F1 (88.66).