You have answers(text) as features and marks to predict.

Here Latent Semantic Analysis(LSA) will help you. First do some text preprocessing on answers by removing stop words, stemming, lemmatizer. Then use tfidf or count vectorizer or word2vec to get the words score(creating feature set).

You will get matrix of numbers(scores), apply LSA (you can use SVD) on that matrix, then apply machine learning model on that feature set and score as your label and get the prediction.

Other features that might be helpful:

1. Counting misspelt words per answer ( you can use PyEnchant)
2. Total word count per answer, average word length per answer, sentence count.

Cheers!!!