AI Professor

Inspiration

I wanted to have a go at understanding NLP concepts, and using transformer models instead of just TFIDF. I thought that this project would be beneficial, as it covered a broad spectrum of NLP topics, from recommender systems, to summarisation and classification. In addition, I also wanted to have hands on experience with Azure. Lastly, I have worked with TensorFlow before and wanted to try using PyTorch just for learning experience.

What it does

I did not have time to complete all the requirements of the project, and have so far only delved into the classification portion.

How we built it

After having an understanding of the broad concepts of NLP, I wanted to challenge myself further and code everything from scratch using Azure's SDK. Once I ensured that the text was cleaned, stemmed and lemmatized, I think wrote a script to run the model on Azure's platform.

Challenges we ran into

I spent a large amount of time trying to understand how to get the code to run in Azure. The other time consuming portion was trying to build a suitable environment for PyTorch to work in Azure.

Accomplishments / What we learned

I realized that PyTorch does not have built in early stopping, and I should have used PyTorch lightning instead. However, I was too late into the project to make the switch. Another learning point would be to try building things iteratively, and trying to debug the code.

What's next for AI Professor

My project is incomplete, and there are things I would actively work on in the coming weeks. 1) Dynamic summarisation using beam search 2) I have downloaded a set of additional news articles, and would want to see how the trained model fairs on this 3) Registering the model in Azure

Built With

• azure
• python