Abstract
Chemical patents are a commonly used channel for disclosing novel compounds and reactions, and hence represent important resources for chemical and pharmaceutical research. Key chemical data in patents is often presented in tables. Both the number and the size of tables can be very large in patent documents. In addition, various types of information can be presented in tables in patents, including spectroscopic and physical data, or pharmacological use and effects of chemicals. Since images of Markush structures and merged cells are commonly used in these tables, their structure also shows substantial variation. This heterogeneity in content and structure of tables in chemical patents makes relevant information difficult to find. We therefore propose a new text mining task of automatically categorising tables in chemical patents based on their contents. Categorisation of tables based on the nature of their content can help to identify tables containing key information, improving the accessibility of information in patents that is highly relevant for new inventions. For developing and evaluating methods for the table classification task, we developed a new dataset, called ChemTables, which consists of 7,886 chemical patent tables with labels of their content type. We introduce this data set in detail. We further establish strong baselines for the table classification task in chemical patents by applying state-of-the-art neural network models developed for natural language processing, including TabNet, ResNet and Table-BERT on ChemTables. The best performing model, Table-BERT, achieves a performance of 88.66 micro F1 score on the table classification task.
Introduction
Chemical patents serve as a primary repository for the disclosure of novel chemical compounds and reactions, often predating publication in scientific journals by several years. For researchers in the pharmaceutical and chemical sectors, these documents are a goldmine of experimental data. However, the most critical information—such as melting points, NMR spectra, and bioactivity—is frequently embedded within complex tables.
Unlike standard web tables, patent tables are characterized by extreme heterogeneity. They are often massive, averaging three times the number of rows found in typical web tables, and frequently incorporate merged cells or embedded images of chemical structures (e.g., Markush structures). This complexity often causes traditional text-mining tools to bypass tabular data entirely, resulting in a significant loss of information. To bridge this gap, researchers have identified a need for automated systems capable of understanding the semantic content of a table before attempting data extraction.
The ChemTables Dataset
To address the challenge of table understanding, the authors introduced ChemTables, a specialized dataset designed for semantic classification.
- Composition: The dataset comprises 7,886 tables extracted from 1,000 XML-formatted patents sourced from the European Patent Office (EPO) and the USPTO.
- Scale: These tables contain approximately 3 million tokens.
- Annotation: Domain experts (chemists) manually annotated the tables based on the Reaxys schema. Categories include SPECT (Spectroscopic data), PHYS (Physical properties), RX (Chemical reactions), and PHARM (Pharmacological data).
- Structural Characteristics: Analysis of the dataset revealed that patent tables are "tall" rather than "wide." While 99% of the tables have fewer than 15 columns, they can extend to over 1,000 rows. This vertical growth emphasizes the importance of column headers as primary semantic anchors.
Methodology
The researchers framed table classification as a supervised learning task, evaluating several architectures:
- Baseline Models: Naïve Bayes and Support Vector Machines (SVM) utilizing TF-IDF features.
- TabNet: A hybrid model using a Bi-LSTM to encode cell content, subsequently treating the table as an image for processing by a Residual Network (ResNet).
- TBResNet: An 18-layer ResNet architecture designed to test whether deep computer vision techniques could effectively capture complex table structures.
- Table-BERT: A transformer-based model that "flattens" 2D tables into 1D text sequences. Two flattening strategies were tested: Linearization (row-by-row concatenation) and Naturalization (converting cells into template-based natural language sentences).
Results and Discussion
The experimental results demonstrated the superiority of deep learning approaches over traditional machine learning.
- Performance: Table-BERT emerged as the top performer, achieving a micro F1 score of 88.70. All neural models outperformed the SVM (84.89 F1) and Naïve Bayes (80.95 F1) baselines.
- Category Consistency: Performance was particularly strong for high-frequency categories like SPECT, where models reached an F1 score of approximately 95.
- The Importance of Table Body: A key finding was that classification accuracy dropped by over 20 F1 points when models were restricted to only the table caption and the first three rows. This indicates that the full body of the table is essential for contextual understanding.
- Flattening Strategies: Interestingly, simple Linearization outperformed Naturalization for Table-BERT. The inconsistent header structures in patent tables likely caused the natural language templates to introduce noise rather than clarity.
- Transfer Learning: Table-BERT’s pre-trained linguistic knowledge provided a significant advantage in "under-represented" classes, allowing the model to generalize effectively even with limited examples.
Conclusion
The introduction of ChemTables and the benchmarking of state-of-the-art NLP models mark a significant step toward unlocking the data trapped in chemical patents. While the current models show high accuracy, a "context gap" remains; some tables require the surrounding patent text to be fully understood.
Future research will focus on multimodal inputs that combine 2D structural awareness with semantic transformers, as well as developing models that incorporate the surrounding textual context of the patent. Ultimately, these advancements will lead to end-to-end systems capable of not just classifying tables, but extracting specific chemical entities and properties for automated drug discovery.
Summary updated