search

castorini / MP-CNN-Torch

Multi-Perspective Convolutional Neural Networks for modeling textual similarity (He et al., EMNLP 2015)
107 stars 58 forks source link

Code to train on Microsoft Research Paraphrase Corpus(MSRP). #1

Closed aironashish closed 8 years ago

aironashish commented 8 years ago

Hi, I was trying to replicate the results of your experiments on MSRP. But, I can't find the required code. Can you please suggest the required changes to handle the classification task. I guess, num_classes and loss function needs to be changed to 2 and hinge loss, respectively.

Also, I was running trainSIC.lua on a dataset with 2 classes(and I made the required changes like changing num_classes = 2 and in predictCombination function val = torch.range(1,2,1)).But, the dev score results in NAN. But, if I run trainSIC without changing the Conv.lua and trainSIC.lua (dataset contains still 2 classes only). I am getting a dev score of .129. Can you please explain the reason for this behavior?

hohoCode commented 8 years ago

Thanks for your interests.

If you want to change the trainSIC.lua file for your own data set and for binary classification. I think it should be easy to do, besides what you have done, I think it is probably needed to pay attention to this part of codes: https://github.com/hohoCode/textSimilarityConvNet/blob/master/Conv.lua#L89-L102

The training data labels should be set properly, please make sure the util/read_data.lua is able to handle your own dataset.

Also, in the prediction function, since your task is on binary classification, I do not think you still need the dot products as what we have done here, you may need to change the codes here as well: https://github.com/hohoCode/textSimilarityConvNet/blob/master/Conv.lua#L142-L148

So overall I think you need to change Conv.lua slightly for your own data.

aironashish commented 8 years ago

Thank you for answering Part one. And second part is solved (changed few more lines).