We used regex to exclude punctuations in the strings and filtered out the stopwords based on the nltk.corpus.
word_match_sharetfidf_word_match_share
Following ChengLong's idea and code: https://github.com/ChenglongChen/Kaggle_CrowdFlower/tree/master/Code/Feat
Count of words
Count of unique words
Ratio of unique words
Count of digits
Count of unique digits
Ratio of unique digits
Count of q1 words in q2
Count of q2 words in q1
Ratio of q1 words in q2
Ratio of q2 words in q1
jaccard_coef between q1 and q2
dice_dist between q1 and q2
cosine similarity of q1&q2 tfidf
cosine similarity of q1&q2 BOW
cosine similarity of q1&q2 SVD tfidf
cosine similarity of q1&q2 SVD BOW
logistic regression -- train AUC scores: 0.804
xgboost -- test AUC scores: 0.872
LSTM/GRU -- following the leaky feature solution of lystdo:
https://www.kaggle.com/lystdo/lb-0-18-lstm-with-glove-and-magic-features