1
$\begingroup$

I used Data augmentation technique on my dataset, to have more data to train. My data is text so the data augmentation technique is based on random insertion of words, random swaps and synonyms replacement.

The algorithm I used performs well in other datasets, but in mine, it gives lower accuracy results comparing to the original experiment. Are there any logical interpretations?

$\endgroup$
1
$\begingroup$

Text data is at the same time:

  • very structured, because swapping only a few words in a sentence can make it complete gibberish,
  • and very flexible, because there are usually many ways to express the same idea in a sentence.

As a consequence, it's very hard to have a text sample which is representative enough of a "population" text, i.e. which covers enough cases for all the possible inputs. But augmentation methods are practically sure to fail, because either they are going to make the text gibberish or just cover minor variations which don't improve the coverage significantly.

That's why a lot of the work in NLP is about experimental design and preprocessing.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.