������ý

We take digital personal assistants for granted these days. Whether it’s looking for the nearest Mexican restaurant, sending a message or just checking the weather, we’re getting pretty comfortable with Siri and Alexa.

But these systems are still limited: they only deal with one task at a time, and more complicated interactions can leave them confused.

Iris, a chatbot system developed by a team at Stanford University, is different. It can handle more complex forms of conversation – and could pave the way for personal assistants that understand how we really speak to one another.

“I would say this is probably among the most complex behaviour I’ve seen from a chatbot to date,” says Ryan Lowe at research lab OpenAI, who was not involved in the work. “Possibly also one of the most useful.”

When we talk, we use all sorts of linguistic tricks and techniques to make ourselves understood. One of the most common is the way we nest sub-conversations within an overarching discussion. You do this, for example, when you answer the question “when shall we meet at the pub?” by asking a further question about when that person finishes work.

Alexa or Siri struggle with such nested conversations unless they have been preprogrammed – or hard-coded – to react to specific examples.

Iris does it by turning language commands into blocks of text that can be flexibly combined with other ones. This design allows every user command (such as “make a reservation”) to be tagged with instructions that tell Iris how it can be stitched together with further commands.

This also narrows the range of other types of command that the tool can act on in the context of the conversational strand. Thus armed, Iris can thread a series of commands together to make sense of them.

Reading the context

Furthermore, Iris understands another conversational quirk called anaphora: a phrase that depends on an earlier part of the conversation, such as saying “he” when you earlier mentioned your brother. Again, the top digital assistants have this ability, but only when hard-coded.

Iris is still a bit limited, which means that for now it’s only being used as a bespoke data science tool. It lacks the natural language ability that Apple, Google and Amazon have baked into their assistants. But in the future, these could integrate Iris’s underlying architecture, providing “a scaffolding of context” for a future generation of chatbots, says Ethan Fast, part of the team behind Iris.

Lowe cautions that – for now – Iris still only understands a relatively small subset of commands, so he’s not yet sure how well it would scale.

But it is able to learn as it goes, too. That’s why Fast is planning to launch a standalone web app in the next few months to let far more users interact with Iris and improve its understanding – and ours. “We hope we can learn much higher-level stuff about how conversations flow,” he says.

Reference: