91av

SMARTPHONES may soon get a lot better at finishing your sentences for you – with the help of words and phrases gleaned from crowdsourcing.

The software packages in today’s phones often struggle with texts and voice commands if a user attempts words or phrases that aren’t included in the phone’s database.

To see if the crowd could help, of Montana Tech in Butte and at the University of St Andrews, UK, called upon workers of the Amazon Mechanical Turk. The plan was to try and improve a predictive system used in Augmented and Alternative Communication (AAC) devices, which help disabled people to communicate by painstakingly typing out words interpreted from their muscle twitches or blinks.

The researchers paid 298 Mechanical Turk workers to imagine phrases they might need if they had motor neuron disease or cerebral palsy. Their responses produced nearly 6000 useful phrases. Next, by trawling through Twitter postings and other social media texts the workers extracted sentences and phrases with similar structures to the initial phrases, expanding the corpus to tens of millions of entries.

The result is a system that needs 11 per cent fewer keystrokes than a standard AAC device. For anyone struggling over every word it is a big improvement, says Kristensson. It should also work when applied to more standard text and speech recognition systems, he says. The work appears in the .

at the University of Sussex, UK, agrees that the system has potential. “Getting adequate quantities of good quality data to build statistical [language] models is one of the most significant challenges in this area,” he says.