Google's Parsey McParseface helps machines understand English almost as well as humans

Understanding human language is easy for humans -- not so for machines. Making them understand human sentences is one of the biggest challenges in today's field of artificial intelligence (AI). 

Now, Google has open sourced its new English language parsing model, called Parsey McParseface (yes, really), which helps a computer understand human language with an amazing degree of accuracy. 

The process of parsing is analyzing a sentence according to grammar rules, and figuring out what it really means. Parsey McParseface is just the parser (Google tells us it had a problem figuring a good name, and then someone came up with that. We see a worrying trend here), a part of an open-source neural network framework called SyntaxNet. 

The system uses neural networks (remember, that's how Google's AI defeated one of the world's top Go players) to figure out how different parts of an English sentence relate to each other. 

The problem with human language is that it's very ambiguous. You and I understand what "Bob crossed the street and entered a car" means, but a machine might encounter problems -- has Bob walked over the street, or has he made it mad? Humans get the meaning fast, because they have a very good idea of what the nouns in this sentence -- Bob, street, car -- can and cannot do, but machines haven't had the advantage of experiencing the world in quite the same way as humans. 

SyntaxNet solves this by figuring out the different ways a sentence could be understood, giving them scores, and then making decisions based on those scores. "Instead of simply taking the first-best decision at each point, multiple partial hypotheses are kept at each step, with hypotheses only being discarded when there are several other higher-ranked hypotheses under consideration," Google's senior staff research scientists Slav Petrov wrote in a blog post Thursday. See GIF below for an example of how this works. 

Original image has been replaced. Credit: Mashable

So how how accurate is Parsey? Google says it achieved over 94% accuracy on one standard benchmark, which is very close to the 96-97% achieved by trained linguists. On a set of random sentences from the web, Parsey achieves "just over 90%" of parse accuracy. So, not perfect yet, but eerily close to human levels. 

The implications of such an accurate parsing engine are far-reaching. It could, for example, mean a virtual assistant such as Google Now will better understand your queries; it could also lead to a far more accurate Google Translate. 

You can read more about the science behind the project in Google's research paper. If you want to implement SyntaxNet and Parsey in your work, go here. 

Have something to add to this story? Share it in the comments.