For awhile from 2015 to 2020 it seemed as if every start-up touted itself to be powered by AI. The buzz around artificial intelligence has faded a bit, probably because it’s clear that AGI is not just around the corner and also because the media is always looking for the next shiny thing to discuss. But for those of us in the trenches thinking about AI research every day, there remains a critical question to ask - is this latest round of AI, namely around deep learning with SGD, merely a feature to sprinkle onto existing tools and services, or is this truly a new foundation on which to build new technologies?
Dialogue systems and chatbots are going through the same cycle of adoption seen in previous technology growth curves. As a quick primer, we note that mobile experienced the same four phases as it has expanded from technical oddity to ubiquitous usage. In particular, in the first phase, you had a limited number of forerunners who used large brick phones. This certainly didn’t live up the promise of mobile, but it was also certainly distinct from its predecessor of the corded phone. In the second phase, there was a shift to enterprise with Palm Pre, Blackberry and other PDAs. In the third phase, we had the original iPhone which lacked an App Store and other key functionality, but at this point you knew mobile was going to take over the world. Finally, in the fourth phase, there was also Android, long-lasting phones with giant screens, and all the bells and whistles we expect today.
When trying to understand user belief, a NLU model attempts to track the intent over the length of the conversation. However, what format should this intent be represented? Is it continuous or discrete? It it directly passed into the Policy Manager or should it be augmented first? Hundreds of hours and effort will be spent finding labels for training such a model, so it seems reasonable we should agree on what format this label should take. But considering the issue in any depth will show trade-offs in different label formats, so the answer is not immediately obvious.
Suppose you had sample = tensor([[3,5,4] [0,2,1]])
Some notes to remember when building intelligent task oriented dialogue agents: