bowlesbe

4 ms·

1. 1.▲
   by bowlesbe 10y ago

   Could you elaborate? I'm not sure if I follow
2. 2.▲
   by bowlesbe 10y ago

   haha, absolutely. It takes a lot of intelligence to detect non-informativeness. you might enjoy: http://journal.sjdm.org/15/15923a/jdm15923a.pdf
3. 3.▲
   by bowlesbe 10y ago

   Are you sure about the padding? On page 1746, bottom right it says "padded as necessary". And intuitively it makes sense that all your inputs need to be the same size for a CNN.
4. 4.▲
   by bowlesbe 10y ago

   This is a great point, would be worth further investigation. And I agree with your general interpretation. It would be interesting to look further at where CNN is failing to detect bad ones and where the feature engineered one picks them up
5. 5.▲
   by bowlesbe 10y ago

   I think deep learning can be seen as a class of machine learning techniques with more flexibility and which uses neural networks (usually with quite a few layers).
6. 6.▲
   by bowlesbe 10y ago

   I appreciate this!
7. 7.▲
   by bowlesbe 10y ago

   This is actually a great point. Thanks for sharing. I should maybe considering removing LIME in that context or changing the wording.
8. 8.▲
   by bowlesbe 10y ago

   Thanks, would you mind expanding? I also played around with some char CNNs. They had similar performance.
9. 9.▲
   by bowlesbe 10y ago

   50K is not that tiny, but I'd say in the hundreds is pretty small. I'd say hundreds of training examples is pretty normal in academia, but typically quite small in industry, particularly for problems that are actually quite abstra
10. 10.▲
    by bowlesbe 10y ago

    I would agree. Its kind of amazing that it indeed it does seem to work better than simpler models.
11. 11.▲
    by bowlesbe 10y ago

    This is really great idea. Actually if there is something you can share along these lines, that would be amazing. I know Crowd Flower has a great "internal only" tool, which is kind of similar to what you are designing, but you h
12. 12.▲
    by bowlesbe 10y ago

    Thanks! I'll check it out. I have also been reading about abstractive summarization - hard problem it seems!
13. 13.▲
    by bowlesbe 10y ago

    Great question and I share your intuition but I think its all properly regularizing your model. I guess for neural networks, Dropout works really darn well as a regularization strategy. I could have tried to see whether performance dropped
14. 14.▲
    by bowlesbe 10y ago

    This is correct. We had 300-400 examples of each
15. 15.▲
    by bowlesbe 10y ago

    Great point! I considering using fasttext as a baseline, however in practice fasttext really didn't work well at all with the small data set, much worse than the tfidf baseline. I think Fasttext's classification approach might not
16. 16.▲
    by bowlesbe 10y ago

    Thanks for the comment! There were a few hundred sentences of each, collected internally from from a wide number of descriptions. Yes, I'd definitely agree- more data is what we need here for further model improvements.