Ubuntu Help AI Chatbot

Demo for interaction with AI chatbot

Inspiration and what it does

We wanted to complete the FINRA Tech Challenge to create an AI chatbot that handles common queries and redirects to a human representative otherwise.

How we built it

We built the project using the Azure QnA Maker API and an Azure Web Bot as a wrapper. We used knowledge bases with previous queries and answers and matched each new query with a certain confidence rating. If this confidence rating was below a certain threshold, we prompted a redirect to a customer service representative.

Challenges we ran into

At first we tried to look into the Maluuba frames dataset and flattening the parsed data json to make a tensor, but the dataset proved much too convoluted and cumbersome. At a different point in time we were trying to implement the wiki-text bot Adam and training multiple models so that it would provide the same output for queries it was confident on, and different output for outlier entries. However, the Natural Language Processing component used hard-coded AI rather than machine learning models, so this was not possible.

What we are proud of and what we learned

We had to scrap our ideas multiple times and did not start implementing the final version until around midnight of the last night, but we learned in detail how online chatbots work, why our previous ideas did not succeed and we also learned a lot about natural language processing. In the final implementation, we learned how to interact with the Microsoft Azure APIs and how to link them together.

What's next for Ubuntu Help AI Chatbot

For this short duration project we used a streamlined version of the Ubuntu Dialogue Corpus knowledge base, but with more time we could find ways to use and manage the whole large dataset, and also implement online learning as new queries are made. In addition, each query is considered by the bot independently, but in the future we could find a way to have consecutive queries interact with each other.