TODO

Things which I believe would be most useful -

1. Create Flask APIs - prediction function should be exposed via flask, so that it becomes implementation agnostic
2. Improve architecture
3. Add error handling and add comments
4. Add higher weightage to classes which fall in the category of expletives, or escalations -

• One way is to have separate models for various categories, for e.g., expletive, escalations and business specific queries.
• Add a voting classifier on top of all models to correctly estimate the response

reason point 4 is required because the bot confuses between queries like "You're Cute" and "You're Stupid", and responds in a similar fashion.

5. Add normalizer to look for spelling mistakes
6. Add lemmatization and stemming to improve accuracy
7. Add additional classifier for estimating category of question. This can be used to send input to selective classifier based on category.
8. Add follow-up queries support.
9. Add process to record unclassified queries

• If system identifies an unclassified query, it should be appended, to be able to access at later stage, for retraining.

10. Restructure the training data storage. More on this below.
11. Move data to database instead of csv file. that ways, it would be better managed.
12. Implement an GUI for adding data/ training, once point 10 is completed.
13. Add Multi-lingual support.