Siri is quite good at recognizing what we say, but used to run into difficulties with the names of small businesses. That was until Apple developers found a way to make her much better at this task.
In fact, the new system is more than 40 percent less likely to come up with the wrong business name.
A post on Apple’s Machine Learning Journal summarizes the problem: “Accurately recognizing named entities, like small local businesses, has remained a performance bottleneck.”
Speech recognition systems learn how words are pronounced by hearing people say them. The names of local businesses aren’t used often enough for Siri to learn to recognize how they sound
And there’s another problem. The words we say sound quite a bit like lots of other words. Siri decides whether a particular set of sounds indicates certain words by how often those words are used. A local business name doesn’t get used very often, so Apple’s computers consider other words to be more probable because they get used more frequently.
Take Shillings, a famous Marietta restaurant. Siri doesn’t have tremendous experience in hearing people ask for directions to it, and so could have problems with that name.
Sort of like Yellow Pages for Siri
Apple solved this problem by building a geolocation-based language model (Geo-LM).
Developers broke the country up into 169 areas that are economically and socially linked, and created databases for each. It then added grapheme-to-phoneme pronunciation guides for the names of the businesses in each of these areas to the local database.
When someone asks their iPhone something like “Give me directions to…” the system knows that what comes next could be a business, so it looks in the local Geo-LM. Words in the database that sound like what the person said are assigned a higher likelihood.
The new system for identifying local businesses isn’t perfect, but Apple promises that it brings significant improvements.
A collection of San Francisco-related test phrases were given to Siri. Without the Geo-LM, the system got the wrong answer 26.5 percent of the time. When the appropriate Geo-LM was added, that dropped to 15.1 percent, a 43 percent improvement. The developers performed similar tests on regions across the U.S., with quite similar results.
This new system isn’t pie-in-the-sky theory. It’s already part of Siri. So we here at Cult of Mac performed our own simple test. We asked Siri directions to Shillings, and to the regional pizza chain Mellow Mushroom. The system had no difficulties with either request.
Now we just need the ongoing improvements to Apple Maps to reach our area.