How to call training and testing on new data?

[lucas0]

In your instructions you say:

to change s1['test']['path'] and s1['unlab']['path'] but there are two other variables (target['test']['path'] and target['test']['path']) that hold the target values depending on the data you're using to train/test your model.

What should we assign to those? From your instructions seems that we should use the target values for twitter/yelp/movies for our dataset.

How have you arrived on these values?

My data2.py and train.py/test.py modifications are like this:

if dat=='my_data':
        s1['test']['path'] = os.path.join(data_path, 'my_data_test.txt')
        s1['unlab']['path'] ='dataset/pdtb/my_data_unlabeled.txt'
        target['test']['path'] = os.path.join(data_path,'twitterl.txt')
        targetv['test']['path'] = 'dataset/pdtb/tlv.txt'

if params.test_data=="my_data":
    _,xst= getFeatures(os.path.join(params.nlipath,'my_data_test.txt'))
    _,xsu = getFeatures('dataset/pdtb/my_data_unlabeled.txt')

And then the calls are:

python train.py --gpu_id 0 --test_data my_data

python test.py --gpu_id 0 --test_data my_data

This is working, but I'm not comfortable using the twitter target['test']['path'] and target['test']['path']. Is there any way I could generate those values for my data?

Thanks a lot for the work and making the code available. This will help me a lot in my work.

Best, Lucas.

[wjko2]

This system is designed to adapt to new domains using only the sentences from the new domain, and does not require specificity labels from the new domain. The target['test']['path'] is used only for evaluation purposes, and does not effect the result, so you don't need to generate them. Those labels were annotated on MTurk.

[lucas0]

Thank you very much. Are those calls the right way to do it for a new domain?

python train.py --gpu_id 0 --test_data my_data

python test.py --gpu_id 0 --test_data my_data

Also, how would the model work for longer sentences? Is it ok to put a paragraph with multiple sentences as a single input? Would the varying length of the paragraphs be an issue?

[wjko2]

Yes, It's the right way. It is ok to use multiple sentences as a single input. I have only run this on the released datasets and some dialog datasets, so I am not sure how it will work if the inputs are very long.