Feature Engineering: Grammatical Features

tldr; Levenshtein distance is the most useful.

Distance measures how different raw sentence from essay is to corrected sentence by T5 model. The model used is T5-GEC-small. The model corrects sentences from the essay. Then distance measures are used to measure how different the sentences are.

Importance-wise Levenshtein distance is more effective than hamming, cosine, and jaccard.

Operation-wise sum and max operations are more important than others.