Fig. 5:

Model prediction accuracy of the four tokenization schemes. We did not include the node-ID-based method in character level accuracy figure because it is vague to define it when the sequence lengths varies too much represented by different tokens.