Apple Reveals How AI Gurus Made Siri An Expert In Local Businesses

Siri is very good at recognizing what we are saying but was having trouble with the names of small businesses. It was until Apple developers found a way to make it much better in this task. In fact, the new system is less than 40% less likely to present the name of the company.

A post on Apple Automatic Learning Log summarizes the problem: “Accurate recognition of named entities, such as small local businesses, has remained a bottleneck in performance.” Voice recognition systems learn how words are pronounced by hearing people say them. The names of local businesses are not used enough for Siri to learn to recognize their sound

And there is another problem. The words we say look a little like a lot of other words. Siri decides if a particular set of sounds indicates certain words by how often are these words used. A local business name is not very often used, so Apple computers consider that other words are more likely as they are used more frequently. Take Shillings, a famous Marietta restaurant. Siri does not have much experience with hearing information requests, which could cause problems with this name.

Sort of Yellow Pages as Siri

Apple solved this problem by creating a geolocation-based language model (Geo-LM).  The developers divided the country into 169 economically and socially related areas and created databases for each. He then added to the local database pronunciation guides from graphene to phoneme for the names of companies in each of these areas.

When someone asks his iPhone something like “Give me instructions for …”, the system knows that what comes next may be a business, so it looks in the local Geo-LM. The words in the database that look like what the person said are more likely to be associated.

It works

The new local business identification system is not perfect, but Apple promises significant improvements. A collection of test sentences related to San Francisco was given to Siri. Without Geo-LM, the system received a wrong response 26.5% of the time. When the appropriate Geo-LM was added, it dropped to 15.1%, a 43% improvement. The developers have performed similar tests on regions across the United States, with fairly similar results.

This new system is not a theory of pie in the sky. It’s already part of Siri. So we here at Worship of mac performed our own simple test. We asked Siri for directions for Shillings, as well as for the Mellow Mushroom regional pizza chain. The system has not encountered any difficulty with one or the other request.

Compsmag