Experimenting with fasttext on tweets. Predict which Tweets are about real disasters and which ones are not Trains on CPU , compatible with Linux cmd line for auto-tune-validation.

• label_prefix : The librarry needs a prefix to be added to classification labels
• lr: The learning rate...works well with default (0.1) lr.
• neg : Number of negative samples 2<neg<6
• epoch : 5,10,15
• dim : 128,256 perform very well.
• loss: softmax takes a bit longer ,hs hierarchial softmax is good too, ns not good score.
• word_ngrams: 2=bigrams ,3=trigrams ,in this case limit it to 2 as per original paper + score_performance
• ws: Size of context window ,here avg sentence length is not too large ,therefore we chose 3 based on experiments.
• bucket: Hash length -**

Above task can be completed using TFIDF,SVM,BERT,GloVE,LSTM, etc but training time is really high and may require strong GPU.

• PythonDocs
• Enriching Word Vectors with Subword Information
• Bag of Tricks for Efficient Text Classification
• Open-sourcing hyperparameter autotuning for fastText
• English word vectors
• Original Paper Presentation