MathWriting+

Overview

MathWriting+ extends the original MathWriting dataset with detailed structural annotations for handwritten equations. It includes labeled symbols, relation graphs in MathML, and refined trace group data across multiple subsets (train, validation, test, synthetic). This enables handwriting recognition approaches that leverage structural information. For more details, see "The Return of Structural Approaches in Handwriting Recognition".

Dataset Breakdown

The table below summarizes the labeled data versus the total available for each subset. (All numbers are rounded for simplicity.)

Subset
Annotated Count
Total Count

Symbols
~6.3K
~6.4K

Train
~143K
~230K

Validation
~9K
~16K

Test
~6K
~7.6K

Synthetic
~86K
~121K

Symbols: Isolated handwritten mathematical symbols

Train, Validation, Test: Handwritten mathematical expressions collected from various sources.

Synthetic: Artificially generated mathematical expressions.

The dataset includes only those samples that have been enhanced with additional structural annotations (trace groups and MathML).

Limitations and Quality Assurance

We have not yet labeled more complex constructs such as matrices, vectors, or symbols with overlines/arrow notations.

The annotations have been cross-checked using a state-of-the-art symbol classification model further described in the corresponding paper. While most of the data is highly accurate, there might be a few errors, especially for symbols not included in the CROHME dataset, so users should take these into account during their research.

Usage and Citation

If you use the MathWriting+ Dataset in your work, please cite this dataset or our paper:

"The Return of Structural Approaches in Handwriting Recognition"

Your citation helps acknowledge the effort behind the dataset and supports further advancements in interpretable handwritten mathematical expression recognition.