This article is from the archive of our partner 
.
Despite hiring a couple of linguists to get its new search engine to move "beyond 'Robospeak" and understand how people talk, Facebook hasn't taught Graph Search how to do that very well just yet. And that's a problem, no matter which way the social network spins it. Unlike Google's pattern-matching search engine, Facebook's new recommendation-based social search platform tries to understand full sentences. And that takes context, something that's very hard to teach even the smartest computers, as one of the linguists that worked on the project, Amy Campbell, told The New York Times's Somini Sengupta.
In order to think more like a person the Graph Search team taught the engine 25 synonyms for "student" so that when someone types in "Stanford Academics that work at Facebook" the engine knows to look for "students" — 275,000 different ways in fact. But it turns out that an English class isn't the future of machine learning: a grammar and vocabulary lesson proves a lot easier than complex sentient thoughts, and that's where Facebook's new product breaks down in practice.
For example, Graph Search doesn't get vague pronouns. My query today for "photos Elle Reeve likes that she commented on" confuses Facebook's beyond-robo engine. Instead, Graph Search results track down photos that my Atlantic Wire colleague "likes" but that I commented on:
The new Facebook engine isn't giving me these other options because those photos that Elle both liked and commented on — well, they don't exist. It took about 30 seconds of searching to find a recent photo that matched those constraints. Yet Graph Search's machine can't understand that I want liked photos with comments because the request is just vague enough. It doesn't have the right context.
But Facebook's ambiguity problem extends beyond "I" and "she." Graph Search also has problems with double entendres, or sentences with nuance. The phrase "sports fans that like Lady Gaga play" has multiple meanings, notes the Times, especially because the word "fan" has its own special meaning on Facebook. (People with "Pages" have "fans.")
It's not impossible to fix these specific issues. Facebook can add more relevant "context" to Graph Search as more people use it (beta testing rolled out over the last week). But never once has a machine perfectly understood our natural language. IBM's Watson has come close, but it still made an embarrassing mistake every so often, and newer robots like Georgia Tech's Simon are still getting there. Messups are okay (and entertaining) for a computer on a gameshow, or robots that might end up really helping bridge the computer-human divide. But, if I'm really going to use Graph Search as a way to find things in my day-to-day life, right now, those kind of hiccups should happen rarely to never — and Facebook's slow phase-in excuse isn't cutting it. If Graph Search can't understand what humans want, it's simply not doing its job.
This article is from the archive of our partner The Wire.
