Data in Action: Helping AI Recognize Intent
Artificial intelligence driven by deep learning technology is gradually working its way into our everyday lives. Products that incorporate some kind of artificial human language understanding are now in our homes, our cars, and even at our banks.
Finn Ai is a software company focused on developing artificially intelligent conversational assistants to help people with their banking needs—providing their customers with a helpful and intuitive online experience, even when there’s no human staff member involved. In the financial domain, mistakes are costly, so it is important that the AI understands exactly what the user intends to do.
As part of their capstone project, students from the University of British Columbia’s Master of Data Science program (MDS) partnered with Finn Ai to investigate whether their existing neural network model for identifying user intent could be improved by looking more closely at the language of the tough cases (where the model fails) and reconsidering the space of possible customer intensions.
They found that in some cases the AI was being confused because its original creators hadn’t fully anticipated how the customers would use their system. Using a mixture of statistical visualization and direct inspection of examples, they discovered that some customer needs weren’t well covered by the model, and others were, in a sense, covered too well, with multiple possible interpretations of a single human request.
Ultimately the project demonstrated how a potent mixture of automated data analysis and targeted human intervention can help get an errant language AI back on track.