rajagurunath / AV_NLP_competition

AV_NLP_competition

I am quite new to the AV competition ,this above code managed to get Rank 30 in the Public leaderboard and Rank 63 in private leaderboard
I enjoyed the whole process of the competition !!!!

Approach Taken:

1. Tfidf + Logistic regression

2. Extarcted features with flashtext ,countvectorizer and used as features along with tfidf + logistic regression +RandomForest +Gradient boosting

3. Used an Universal sentence embedding from tensorflow_hub via tensorflow as features + logistic regression+RandomForest +Gradient Boosting +lightgbm +xgbost

4. Ensembeling all the above models result and trained a Logistic regression on the training data and extracted the predictions for test data for all the above models and gave final predictions

5. Using spacy CNN clssifier for text classification : (got highscore among other models -due to the usage of pretrained models in the spacy)

6. Extracting the wordvec vectors for each word and summing those words for a given sentences used as features for above specified Algorithms(models)

7.Final score model : -using Spacy sentence vectors (mean embedings of word2vec) along with universal embedings from Tensorflow +handcoded features with flashtext + countvectorizer (such as using separated columns for vulgar words used in the tweets etc) totally got nearly 918 features approx. 8. For above built features applied a Multi-layer Perceptron from Torch along with SKORCH for easily fiting the classifier .

Future Plain (tried to implement) :

* ULMFIT from Fastai for text classification 
* ELMO