Facebook develops semantic search

Facebook has been playing around with search since it recoded the entire platform and gave us Timeline back in 2012. It is an indisputable fact that Facebook is sitting atop a motherload of information willingly supplied by Facebook users, but mining it adequately has always been problematic for the social network.

When Facebook launched Graph Search a year after Timeline it was hoped that many of the problems were solved. They weren’t but there was some hope in the horizon. The extension of Facebook’s Graph search to now also include mobile seeks to broaden the scope and raise the stakes even higher. 

Because Facebook is as large as it is and because it will undoubtedly continue to refine its search marketing efforts it’s worth taking a closer look at what it has done and why. It is necessary here to break down (and oversimplify) the elements that make up a good search engine. 

In order of importance they are:

  • The Interface
  • The Index
  • The Technology that powers all this (the algorithms) 

Ideally you want the interface to not just be user friendly but, like Google’s search box, to provide feedback data that helps refine and improve the end-user experience. Google’s Hummingbird takes this into account in a constantly improving iteration that leads to an ever more relevant search experience. 

You also want the index to be fast which means its structure has to have some way of navigating through the information it contains. Speed, in an indexing database comes down to computing power in relation to the size of the data subset it has to trawl through and that basically means algorithm parameters. In other words if you program your algorithm to look deep and go beyond the surface values it will require more time to ascertain some of the relationships. 

Finally you need programming parameters that make sense of it all. This is where the quality of the search algorithm really shows because it is key to achieving an acceptable benchmark that can then be improved by everything else. 

Despite the simplicity of the parts there is an inherent complexity to all this that makes search hard to get right. Microsoft knows that. It threw over $5.5 billion at Bing from its launch date in 2009 without ever managing to make it get much traction (in the US it rose to 29% of the market when it combined with Yahoo, which it powered). 

So, How is Facebook Doing With Search?

Facebook has an inherent problem when it comes to computing power: it is expensive. Unlike Google who can throw computing power at search Facebook requires its computers to power and sustain Facebook and then also do search. No resource is ever unlimited and it is no exception here. To economize Facebook has taken a few shortcuts: 

  • Keywords – although Graph Search does not use keywords to deliver results Facebook uses keywords internally to map searches to objects, places and people. There is a problem with that type of mapping in that it makes disambiguation difficult unless you take into account a lot more relational values which eat up computing power. To address the problem Facebook extended, Unicorn, its internal index database (which was used by Facebook engineers to look up entities) to become a search engine.
  • Information retrieval framework - In order to make Unicorn responsive to searches without making it expensive in computing power Facebook throttles searches by applying various constraints on attributes in relation to the entities being searched for. In plain English, if you are looking for something specific Facebook will return values from your friends and acquaintances based on the calculated strength of your connection rather than from its public stream where information has been shared by a stranger, even if that stranger is perfectly OK with the information being publicly available.
  • Intent – Despite the fact that all searches take place within the Facebook environment, Facebook ignores the searcher’s intent. Again this is a saving in computing power (and perhaps algorithm engineering). 

Some of the shortcomings of Graph Search in relation to Google search were mapped out by Mark Traphagen in a brief post on Google+. 

Is Facebook Search Useless Then?

The fact that we are talking about this shows that Facebook’s search still holds potential. More specifically, where local searches are concerned Facebook is brilliant at surfacing recommendations like Yelp’s sourced directly from friend posts or responses to requests put on Facebook by friends. Search queries like “What’s the best dentist in my area?”, “How do I find the best restaurant in Manhattan?” show the power of Facebook’s social Graph to bring up content that few other search verticals can surface successfully. This has a strong marketing implication we will get back to in a minute. 

Unicorn, at the moment, however runs the risk of creating a real search bubble. With the only results being surfaced being from those friends and contacts you have a strong interaction with the Facebook algorithm essentially whittles down your search funnel to an ever narrowing horizon based upon your previous activity which provides a smaller and smaller subset. 

While this may save computing power and (perhaps) increase accuracy. It also locks Facebook users into an ever narrowing interaction box from which the only way out is through adverts. The cynics will smirk here and suggest Facebook is doing it again: making advertisers pay and forcing users to click on ads in order to find what they want. 

I am not sure the ploy is as cut and dried as that. Without the kind of computing power that Google has, Facebook has to prioritise what it shows to its users. The close-friend network approach makes sense in that context and an injection of quality content from the Facebook News Feed to broaden the scope and stop things from getting claustrophobic makes sense. This is also what Zuckerberg advocated in his first public Town Hall Q&A

Facebook’s Semantic Search Approach

Without the ability to truly account for the searcher’s intent and with disambiguation being a little iffy, Facebook needed a way to drill down to the right answer to a search query from the narrow subset of close friend connections. 

Here its semantic search works really well. Facebook uses a high level of natural language processing to differentiate between ambiguities and increase accuracy in results. Its semantic parsing of complex sentences (a requirement to make Facebook Graph Search really work) provides a perfect example of semantic search analysis of search queries. 

The figures below show the paths, both conceptual and practical that deliver results for the search query “My friends who live in San Francisco, California”. 

Fig 1:

Facebook Graph Search

Fig 2:

Facebook Graph Search Semantic Analysis

The Pay-Off For Marketers

Despite the narrow constraints of Facebook’s semantic search marketers now have real opportunities and, the best news is that you do not even have to be a member of Facebook in order to take advantage of this. 

If you have been following the trends since semantic search was implemented and:

  • Have made the quality of the end-user experience paramount.
  • Regularly produce top-quality content.
  • Have been building relationships with online visitors.


Then you will have the kind of online reputation that leads to finding your own evangelists. People who are willing to share your content and discuss its merits, customers who are willing to talk highly about you to their close circle. 

Quality, quality, quality are the three things you need to focus on in order to break into Facebook’s semantic search (and that includes mobile). You do not really need to have a presence there (unless you really want to) and you do not need to buy Facebook advertising (unless, again, you really want to). 


The Infrastructure of Graph Search (Facebook)
Natural Language Interface (Facebook)
Facebook’s Top Open Data Problems
Facebook’s Graph Search Comes to Mobile
Facebook’s Creepy Search Engine (The Register)