A whopping majority — over 80% — of Internet users type search phrases into search engines to find information. But now, Facebook is hoping to enter into the world of search engines with the implementation of their DeepText.
Described as a deep-learning based text understanding engine, the goal of Facebook’s DeepText is to provide almost human accuracy to the process of sorting through all the social media posts on its platform.
This system is able to understand the textual accuracy of several thousand post per second, in over 20 different languages. The company also hopes to expand the system’s natural language processing (NLP) abilities, as these are typically confused by slang.
That means Facebook is developing a search engine that will understand acronyms and phrases like “LOL,” “OMG,” “BTW,” and “bro.” This word-embedding software is also able to understand and compare similar semantics across different languages. For example, Facebook will learn to recognize “feliz cumpleanos” and “happy birthday” as having the same meaning.
As reported on Enterprise Tech, the Facebook researchers explained the importance of this technology in a statement posted on their website:
“Text understanding on Facebook requires solving tricky scaling and language challenges where traditional NLP techniques are not effective. Using deep learning, we are able to understand text better across multiple languages and use labeled data much more efficiently than traditional NLP techniques.”
DeepText also promotes text understanding, which will help filter out unrelated news stories on the Facebook News section of the site. It will also determine a general classification of what the search is about and then provide relevant information that the user could be interested in.
Take the word “basketball.” If a user was to search for a particular basketball team on Facebook’s search engine, DeepText would recognize the the search terms and produce names of players, stats from games, and other information on the teams searched.
In addition, Facebook also hopes to create a software that captures contextual dependencies between words. They have been successful so far, with an error rate of less than 20%.