EmotionAnalyser Team 12

Inspiration

We are often worried about the tone of messages that we send to other people, hence we created an application that is able to determine how our messages are being perceived by others.

What it does

it analyses a message and determines how positive and negative it is to give you the right level of emotion you want.

How we built it

For the chatbot machine learning logic trained a BERT model using the the PyTorch transformers library and the fine-grained 5-label SST dataset.

Challenges we ran into

Spent a huge deal of time deciding on what dataset to use, what multi-text classification architecture to use, what training environment to stick to local vs google colab.

Having a scarce understanding of the the transformer library run into a lot of technical debt and did not manage to use the saved model and connect it with the server where the chatbot application is running.

Accomplishments that we're proud of

Having little experience with NLP, it feels good to now be able to implement a seq2seq model for a tex-classification task.

What we learned

Spend a huge deal of time during the first day doing literature review on the existing deep learning techniques for sequence modelling which proved to be disadvantage

What's next for EmotionAnalyser

Integrating the backend with the NodeJS application by possibly deploying the model on a Flask server and accessing the model evaluation script with a RESTful API. Further training the transformer or potentially trying to achieve a better accuracy just by training on the binary SST dataset.