Armchair football fans are often damning about the broadcasters' commentators. “Any fool could do that job,” is a common complaint. Any fool, perhaps, but what about a computer? Could mindless automatons do justice to the beautiful game?
A group of Disney researchers have set out to achieve just that: building an automatic commentary generation system, capable of describing a football game's action. Ultimately, the group want to demonstrate just how well computers can be taught to use human language – making it a novel twist on the Turing Test. Is it possible to create a computer commentator that could ever rival humans?
Hannaneh Hajishirzi, and her colleagues at the Disney Research, Pittsburgh labs, which are affiliated with Carnegie Mellon University, have just published details of their on-going efforts to create a RoboMotty, a system that could generate plausible commentary to fit alongside the action on the pitch.
To achieve this, the team set out to examine whether they could create a system that enabled a computer to listen to broadcast commentary, and understand what had just happened.
They studied footage from eight English Premiership matches, and how the commentators described the games' events.
“We formulate the problem of understanding soccer commentaries as learning to align sentences in commentaries to a list of events in the corresponding soccer game,” the researchers noted.
So when the system is given a block of commentary, it identifies the events associated with the commentators' descriptions.
For example, when given the following snippet: “Chelsea looking for penalty as Malouda’s header hits Koscielny, not a chance as it hit him in the stomach,” the system can understand the action involved passes, headers and potential fouls. The system is also taught to identify players' names, and this which team they play for.
By analysing the rich variety of language to describe the games' events – goals, fouls, passes etc – the team are creating the building blocks of a commentators lexicon.
Ultimately, the team hope their work will go some way towards improving the interactions between human and computers, by making computers better at understanding how to use our language.
The team still have plenty of work to do: it's hard for their system to interpret commentary in instances where multiple events happen in short succession; pundits' off-topic ramblings are as incompressible to the computers as they are to many listeners.
But once they have established how commentator's talk, it should be much easier to build a computer mimic, the researchers reckon.
Cotton seedling freezes to death as Chang'e-4 shuts down for the Moon's 14-day lunar night
Fortnite easily out-earns PUBG, Assassin's Creed Odyssey and Red Dead Redemption 2 in 2018
Meteor showers as a service will be visible for about 100 kilometres in all directions
Saturn's rings only formed in the past 100 million years, suggests analysis of Cassini space probe data
New findings contradict conventional belief that Saturn's rings were formed along with the planet about 4.5 billion years ago