uda VS nlpaug

Another thought is to do some data augmentation using back-translation, a la https://arxiv.org/abs/1904.12848

The words that replaces the original word are chosen by calculating TF-IDF scores of words over the whole document and taking the lowest ones. You can refer to the code implementation for this in the original paper here.

You'd want to use an NLP method for this as in order to determine optimal homonyms there would have to be some method of deriving context from the words ahead of and behind the substitution. Take a look at nlpaug.

3) Use deep contextual models with wordpiecing/BPE tokenizers- like all the models: BERT, RoBERTA, etc. On the simpler side, could also swap words with synonyms, which is easy to do with this library: https://github.com/makcedward/nlpaug. Instead of a single n-gram per topic, it might be nice to have a bundle of related words- you could play around with wordnet and see if that's helpful- also easy to do w/ nlpaug.

In practice, even swapping words with dictionary synonyms is a problem because context isn't considered. Lexically sensitive contextual augmentation has become more popular in the last year or two - basically you mask a token using a large language model and then use the model to predict it so it has the full context. It's imperfect, but it's surprisingly useful when you want to upsample data. Nlpaug has an easy-to-use implementation https://github.com/makcedward/nlpaug

A cool library I recently came across for text augmentation is nlpaug, it does a different thing to your approach, but I think both are useful :)

This is a nice starting point: https://github.com/makcedward/nlpaug

Since the dataset is relatively tiny, we are working on augmenting it with NLPAug. We use 2 strategies. Synonymisation and back translation.

Thank you!

The medium post is amazingly written! I basically did the same thing - and you beat me with the data augmentation piece. I tried using nlpaug [0] but it didn't improve the model performance. I'll definitely try swapping book titles around.

[0] https://github.com/makcedward/nlpaug

Are there any shortfalls in nlpaug which justified another project?

Spelling error injection In this method, we add spelling errors to some random word in the sentence. These spelling errors can be added programmatically or using a mapping of common spelling errors such as this list for English.