A blog post published on the Facebook Code subsite stresses the enormous importance of text-based communication to the social media giant. The company’s ability to recommend relevant content and filter out spam will reportedly be vastly improved thanks to the abilities of DeepText.
Using a complex system of neural networks and deep-learning techniques, DeepText is able to understand the textual content of posts written in more than 20 languages at a rate of several thousand posts per second.
Poring over Facebook posts might not seem like the most challenging task, but difficulties arise for nonhuman readers when context comes into play. Previous attempts to create a system like DeepText were stymied by confusing elements like ambiguous terminology and slang.
Traditional methods of “teaching” a machine to read centered around the process of assigning each word to a numerical ID that’s easily recognized by a computer. However, this strategy requires words to be written exactly as in the teaching materials if the text is to be understood.
Conversely, DeepText uses a system that preserves semantic links between different terms. This means that words like ‘brother’ and ‘bro’ are connected, as well as different representations of the same term in different languages.
Facebook is an enormous service, and the company faces a big challenge in parsing all the text that makes its way onto the service and coming up with something that’s relevant. The DeepText system could be a huge help toward separating the wheat from the chaff.
In fact, DeepText is already being implemented in various Facebook services as a test of its capabilities. For instance, the system has been integrated into Messenger in such a way that it can determine when a user is looking to call a cab and offer up a means of doing so — but it will ignore any usage of words like “ride” that don’t refer to a taxi journey.
Development of DeepText is set to continue in collaboration with the Facebook AI Research group.
