Stephan Raaijmakers: 'Humans and systems have to learn to understand each other better'
You can ask virtual assistant Siri about the weather, but you can’t have a real conversation with it yet. You can’t refer to anything that’s been said before, or ask the system why it says what it says. Stephan Raaijmakers, Professor by Special Appointment from TNO, hopes to change this.
Raaijmakers: ‘AI is getting better at global language analysis: what a text is about, whether the tone is positive, or whether J.K. Rowling wrote it under a pen name. To help AI understand language better, I’m working on aligning linguistics, cognition, and AI tools. This raises questions like: How much must an artificial neural network actually know about linguistic structure to understand language? And what role do attention, memory, and reward play in language analysis?’
Understanding leads to trust
Once AI systems communicate better, people will trust them more, says Raaijmakers. ‘Advanced AI systems can, for instance, be trained to recommend cancer treatments. Users of these systems, like doctors, only trust a system once it can explain how it reached its conclusion. This is difficult when the answer involves a neural network with millions of parameters. Ultimately, you want to have a meaningful dialogue with an AI system, and even get it to it change its mind.’
Learning language like a child
And yet, a lot of wonderful things are already possible. TNO, Raaijmakers’ employer, implemented the PAL robot, which uses endless patience and natural language communication to teach children suffering from diabetes to cope with their disease. ‘Systems like these learn language the way we learn a second language, via explicit instruction and a lot of pre-analysed data. Ultimately, I’d like AI systems to learn language and dialogue the way young children learn: incrementally and interactively. This could lead to AI with a more robust language capacity, and new insights into human language acquisition.’
With lectures and practicals, Raaijmakers hopes to get linguistics and computer science students interested in this new method for language acquisition in artificial systems. ‘So the field can progress further. Practice is very important in this context: you have to do AI to really understand it.’