Adding WNUT_2020_NER dataset support

[aynetdia]

This implementation appropriately prepares and loads the aforementioned dataset.

I could also add commit the testing function for the pipenv testing, if needed.

Edit: typo

[alanakbik]

@aynetdia thanks for adding this! There is an import statement missing - can you add it?

Also, I notice the statistics of the corpus are slightly off. If I print the WNUT_2020_NER object, I get:

Corpus: 8075 train + 2740 dev + 2691 test sentences

But the task website at https://github.com/jeniyat/WNUT_2020_NER lists the following statistics:

• train_data: 370 protocols with 8444 sentences
• dev_data: 122 protocols with 2839 sentences
• test_data: 123 protocols with 2813 sentences

Could you double-check?

[aynetdia]

I was able to fix the corpus length problem, however only for the train_data and test_data. Now they contain 8444 and 2813 sentences respectively. The dev_data contains 2862 sentences instead of 2839 as specified in the GitHub repo.

Not sure what the reason for the discrepancy - I went over the whole dev sample manually and couldn't notice any obvious irregularities.

Edit: typo

[alanakbik]

@aynetdia thanks for updating the PR!

Just one more item: There is a folder for the 2020 version of the test data (called test_data_2020), which contains a lot more sentences. Their paper (see Table 1) distinguishes between Text-18 and Test-20 datasets. Test-20 is what was used to evaluate and has 3562 sentences. Can you change it so that test_data_2020 is used instead of test_data?

[alanakbik]

@aynetdia thanks for adding this!