Artificial Intelligence (“AI”) researchers from Stanford University and Facebook appear to be joining forces to develop “self-learning” chatbots.
Traditionally, most of a chatbot’s training happens before it is deployed. But according to a new paper, “Learning from Dialogue After Deployment: Feed Yourself, Chatbot”, scientists from Facebook and Stanford University believe digital assistants of the future will be able to self-improve and learn on the go after going live.
The plan is that these bots, in addition to being trained on the primary task, will be able to predict the user’s satisfaction with its responses. The bot’s ability then improves by imitating human responses when the human user is satisfied, or by asking for feedback when they are not - learning from both scenarios, automatically.
Michael Lovegrove, CEO of New Zealand AI company, JRNY, believes a self-learning chatbot is valuable in theory: “The concept of a digital assistant that learns by itself overtime is fantastic. A chatbot, or digital assistant, that gets better and better only improves the customer experience further, while adding more value and return on investment for companies using digital assistants as part of their sales and marketing strategy.”
JRNY’s Lead Engineer, Adam Kent, adds: “It appears the scientists from Facebook and Stanford are using clever reinforcement learning instead of general intelligence to avoid situations where tech like this has gone wrong in the past. Take Microsoft’s recent blunder with their bot “Tay” for instance. Tay became racist by using totally autonomous, general intelligence. If these reinforcements like Facebook and Stanford have suggested are possible, it would mean bots can indeed self-learn totally autonomously while minimising the risk of another ‘Tay debacle’.”
JRNY is already exploring how it can improve the customer experience with their digital assistants. One of the ways JRNY’s digital assistants improve over time is by following up with a user to ask if the assistant’s response was useful and accurate. If the answer is yes, and the assistant wasn’t expecting that particular response from the user, a positive connection is made for future interactions. If the response is no, a suggestion can be collected from the user and used for future improvements.
Another way JRNY is improving the chatbot experience like this is by using confidence levels. A digital assistant can offer a customer options based on how confident it is that it has the right response for the customer query. If the user selects one of the options, the connection between the input and the response is made and the assistant’s future responses can then be refined.
Although it seems that totally autonomous, self-learning chatbots are on the horizon according to Stanford University and Facebook, the technology isn’t perfect...yet. There are, however, already sophisticated digital assistants being developed right here in New Zealand.