Add Yelp Reviews Datasets

[goodhamgupta]

Hi everyone,

Thanks for this great library! This PR aims to add training and test files for the Yelp Reviews datasets(#11 and #16). Based on the huggingface datasets library, the two datasets added are:

• Yelp Full Reviews Dataset
• Yelp Polarity Reviews Dataset

For the Yelp Full Reviews dataset:

• The train dataset consists of 650k records and a total of 5 labels: 1,2,3,4 and 5
• The test dataset consists of 50k records with the same number of labels as above.
• It can be queried as follows:

%{review: inputs, rating: targets} = Scidata.YelpFullReviews.download
%{review: test_inputs, rating: test_targets} = Scidata.YelpFullReviews.download_test

Similarly, for the Yelp Polarity dataset:

• The train dataset consists of 560k records with a total of 2 labels: 0(negative) and 1(positive)

  • NOTE: The labels in the dataset are present a "1" and "2", with "1" including all reviews with 1 and 2 star rating, and "2" including all reviews with 3 and 4 star rating. (Converted based on huggingface implementation here)

• The test dataset consists of 38k records with the same number of labels as above.

• It can be queried as follows:

%{review: inputs, sentiment: targets} = Scidata.YelpPolarityReviews.download
%{review: test_inputs, sentiment: test_targets} = Scidata.YelpPolarityReviews.download_test

Also, for downloading the datasets, I'm currently using the URLs provided by Fast AI. Do let me know if you would like me to change them.

Thanks!

[josevalim]

Can you please use nimble_csv for CSV parsing? :)

[goodhamgupta]

Thanks for the suggestion! I've made the required change.

[josevalim]

Suggested change

	records
	|> Enum.map(fn x ->
	x
	|> List.first()
	|> case do
	"1" -> 0
	"2" -> 1
	end
	end)
	Enum.map(records, fn
	["1" | _] -> 0
	["2" | _] -> 1
	end)

[goodhamgupta]

This is indeed much so much cleaner! I've made this change.

[josevalim]

NimbleCSV supports full text parsing, so you should consider using that instead, as it is more efficient (especially since you already have the whole binary in memory anyway).

If you want to stream, then it is best to do it from file, using File.stream or similar. :)

[goodhamgupta]

Thanks to your suggestion, replacing CSV with Nimble CSV helped me remove this function completely 😄

Yes I had aimed to use File.stream initially, but I wasn't sure if I should be adding a function specific to a single dataset to the Utils file, which was why I adopted this method instead 😅 I've made the change now.

[josevalim]

Thank you for the PR @goodhamgupta, I have dropped some initial comments.

[goodhamgupta]

Thank you for the kind review @josevalim! I've made all the requested changes.

[josevalim]

LGTM! @t-rutten, your call. :)

[t-rutten]

We really appreciate your PR @goodhamgupta! It looks great :) I just suggested tweaks to documentation.

[goodhamgupta]

Thanks so much for your kind review @t-rutten! 😄 I've made all the requested changes.