Used Siamese network for parphrase classification.paper
Used 100 dimension pre-trained GloVe word-vector provided by Stanford.As main building block of NN used Bidirectional RNN with LSTM as memory unit since sequence length is 50 there might be possibility of longer dependencies.
Used RMSprop as optimiser and default parameters as mention in keras documentation.RMSprop is good choice for training RNN.Strategy for callback is validation accuracy improvement.
As metric for evaluation i use loss and accuracy. I got 0.18, 0.74 as loss and accuracy respectively on Test-set.
Use of TF-IDF in two ways.
- Concatenated output of NN1 and TF-IDF(200 dim) vector and apply 128 dim hidden layer and same goes for NN2 and finally apply softmax layer.
- There might be possibility of Question having length higher than 50. To deal with problem we use TF-IDF to collect most important 50 words than apply model.
Hyperparameter Tuning
Keras, NLTK