## Evaluation Metrics
We will provide a test set for each subtask and the participants are invited to submit the generation results to evaluate their models' performance. Due to Kaggle limitation, we don't use **CER** metric as mentioned on our paper, but instead we use a proportional metric namely **LevenshteinMean** as the evaluation metric. In detail, the **LevenshteinMean** is calculated by taking the average count of the number of character substitution, insertion, and deletion from the prediction compared to the gold label, while the actual **CER** is calculated by taking the total  count of the number of character substitution, insertion, and deletion and dividing them by the total number of characters.

## Submission Format
The participants are invited to submit a CSV file named with "TEAM_NAME.csv" for evaluation. The format of the CSV file should be the same as "sample_submission.csv" in our released data.