Facebook translates natural language interface under Graph Search

Facebook's latest crash course on Graph Search focuses on the natural language interface for deciphering user queries.


Facebook continues to reveal more details about the underlying infrastructure of Graph Search, this time with a deep dive on the natural language interface.

It's a fairly in-depth piece, and there are plenty of nitty-gritty details available over on the Facebook Engineering blog today.

But the gist of the crash course is to offer more insight into how the infrastructure deciphers queries written out in "natural language" (i.e. colloquial, everyday speech -- nothing fancy or even necessarily grammatically correct) and then spits the best possible results back out.

Facebook engineering manager Xiao Li and research scientist Maxime Boucher emphasized in the blog post on Monday that natural language is the best input for Graph Search because it supports more precise queries across the structured data of the Facebook graph.

A natural language program needs to: understand the key words or entities, understand the grammar that combines them, and then translate into the programming language that our Unicorn index understands, and it needs to do all this very quickly.


For example, they explained that by building in certain grammar rules (i.e. changing from "work" to "worked") changes the meaning and should be included, where as changing "photo" to "photos" does not change the meaning and should be ignored.

The engineers added that Graph Search must also interpret different ways of typing the same question, even comprehending misspelled or grammatically incorrect queries as well as synonyms that people use to talk about the same things. (For example, "pics of my besties" = "photos of my friends")

Once again, the Facebook engineering team acknowledged that Graph Search is still very much in its infancy, with only "one percent" of the work complete.

Part of the slow start, according to the team, was attributed to "little real user data" available to optimize the system, which might baffle some considering Facebook is the largest social network on the planet with data for over one billion people.

Nevertheless, Facebook's team built this data-driven infrastructure from scratch, and we'll be hearing a lot more about it over the course of the year.

Charts via The Facebook Engineering Blog

More about Facebook's Graph Search on ZDNet:

Show Comments