Tokenization Repair in the Presence of Spelling Errors

CoNLL 2021: Reproducibility Material

Web demo

Try our tokenization repair methods in the interactive web demo.

Web demo

Evaluation web applications

Click through our benchmarks and get a visualisation of the results in the evaluation web app.

Tokenization Repair Evaluation

Spelling Correction Evaluation


The data contains our benchmarks described in the paper, as well as trained models and predicted sequences from all our methods (1GB compressed). In addition, you can download our training data (6GB compressed).

Download data

Download training data

Download training data with synthetic OCR and spelling errors

Corrected ACL anthology corpus

A whitespace-corrected version of the ACL anthology corpus will be made available for download with the publication of our paper. In the meantime, you can explore the corrected corpus below.

Corrected ACL anthology corpus


The code comes with a Docker setup for easy reproducibility. A readme file in the code directory explains how to setup the Docker container. If you are not familiar with Docker, please visit

The Docker container allows you to try our methods interactively, run them on our benchmarks (or on yours!), and run the evaluation. Make targets simplify the program calls and give further explanations.

The latest version is 1.2.1 (June 14, 2021).

Download code

arXiv 2020 paper

Here you find the reproducibility material of our arXiv 2020 paper, including benchmarks, results and trained models: Download arXiv 2020 data