Siamese-LSTM

Using MaLSTM model(Siamese networks + LSTM with Manhattan distance) to detect semantic similarity between question pairs. Training dataset used is a subset of the original Quora Question Pairs Dataset(~363K pairs used).

It is Keras implementation based on Original Paper(PDF) and Excellent Medium Article.

Prerequisite

• Paper, Articles
  • Siamese Recurrent Architectures for Learning Sentence Similarity
  • How to predict Quora Question Pairs using Siamese Manhattan LSTM
• Data
  • GoogleNews-vectors-negative300.bin.gz
  • Kaggle's Quora Question Pairs Dataset
• References
  • aditya1503/Siamese-LSTM Original author's GitHub
  • dhwajraj/deep-siamese-text-similarity TensorFlow based implementation

Kaggle's test.csv is too big, so I had extracted only the top 20 questions and created a file called test-20.csv and It is used in the predict.py.

You should put all data files to ./data directory.

How to Run

Training

$ python3 train.py

Predicting

It uses test-20.csv file mentioned above.

$ python3 predict.py

The Results

I have tried with various parameters such as number of hidden states of LSTM cell, activation function of LSTM cell and repeated count of epochs. I have used NVIDIA Tesla P40 GPU x 2 for training and 10% data was used as the validation set(batch size=1024*2). As a result, I have reached about 82.29% accuracy after 50 epochs about 10 mins later.

Epoch 50/50
363861/363861 [==============================] - 12s 33us/step - loss: 0.1172 - acc: 0.8486 - val_loss: 0.1315 - val_acc: 0.8229
Training time finished.
50 epochs in       601.24