This code is an independent implementation of Learning to Rank Short Text Pairs with Convolutional Deep Neural Networks SIGIR'15 in Keras.

1. Evaluation done using TREC-Eval and without word overlap features.
2. Trained the model for 20 epochs, setting the same parameters as mentioned in the paper.
3. Used only a single dropout layer, after the Dense layer.
4. Used sigmoid as the activation function, instead of softmax (mentioned in the paper).

For the provided data jacana-qa-naacl2013-..

Reported in paper for qg-emnlp07-data

1. Neural architecture as given in the paper, complete with regularization and adadelta, with optimal hyperparameters.
2. Used parse.py from the implementations in Other Implementations section.
3. Implemented map_score() using sklearn.

1. Accuracy printed while training does not have any relevance.
2. Tensorflow(not compiled to use SS4.1 et. al) is faster than theano on same hardware (77m vs 134m CPU time on E5-2680). But, it gave lower scores over the same epochs. Need to dump optimal weights and test on that.
3. And almost identical timings on i7-5820K with TF compiled to use low level optimizations. Theano didn't use lblas here.

1. sklearn.metrics.average_precision_score() gives a division by zero error and is thus unable to compute the scores, return nan.
2. Add early stopping and store parameters with best MAP score on dev set (as per the paper). This will increase the scores a fair bit (The last iteration had a high loss, as compared to some of the previous ones.).
3. Add dropout at a few other places.
4. Use word overlap features.

1. https://github.com/shashankg7/Keras-CNN-QA
2. https://github.com/aseveryn/deep-qa