Add community notebook for T5 sentiment span extraction

[enzoampil]

This is an example notebook that aims to increase the coverage of T5 fine-tuning examples to address #4426 .

This notebook presents a high level overview of T5, its significance for the future of NLP in practice, and a thoroughly commented tutorial on how to fine-tune T5 for sentiment span extraction with an extractive Q&A format.

I recently presented this in a webinar published on youtube.

[codecov-commenter]

Codecov Report

Merging #4700 into master will decrease coverage by 1.41%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #4700      +/-   ##
==========================================
- Coverage   77.14%   75.72%   -1.42%     
==========================================
  Files         128      128              
  Lines       21070    21070              
==========================================
- Hits        16255    15956     -299     
- Misses       4815     5114     +299     

Impacted Files	Coverage Δ
src/transformers/modeling_longformer.py	18.70% <0.00%> (-74.83%)	⬇️
src/transformers/configuration_longformer.py	76.92% <0.00%> (-23.08%)	⬇️
src/transformers/modeling_t5.py	77.18% <0.00%> (-6.35%)	⬇️
src/transformers/file_utils.py	73.44% <0.00%> (-0.42%)	⬇️
src/transformers/modeling_bert.py	86.77% <0.00%> (-0.19%)	⬇️
src/transformers/modeling_utils.py	90.17% <0.00%> (+0.23%)	⬆️
src/transformers/modeling_openai.py	81.78% <0.00%> (+1.37%)	⬆️
src/transformers/modeling_gpt2.py	86.21% <0.00%> (+14.10%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0866669...34c9a46. Read the comment docs.

[LysandreJik]

Nice webinar, and cool notebook :)

@patrickvonplaten, do you want to take a look?

[enzoampil]

Thanks @LysandreJik ! 😄

[patrickvonplaten]

Awesome notebook @enzoampil!

LGTM for merge!

Which dataset do you use exactly to fine-tune T5 here?

[enzoampil]

Thanks @patrickvonplaten ! 😄

For the dataset, I got it from an ongoing Kaggle competition called Tweet Sentiment Extraction.

The objective is to extract the span from a tweet that indicates its sentiment

Example input:

sentiment: negative
tweet:  How did we just get paid and still be broke as hell?! No shopping spree for me today

Example output:

broke as hell?!

[enzoampil]

I was thinking about contributing this to the nlp library, but I'm not sure if Kaggle has policies regarding uploading their datasets to other public sources ...

[patrickvonplaten]

I see! Yeah no worries - I don't think we currently handle dataset processing from on-going kaggle competition links.