Facebook sees text a prevalent of communication on its platform. Using Artificial Intelligence (AI) called DeepText, the social giant is trying to understand the various ways users use text to help it improve experience.
Introduced on June 1st, 2016, DeepText is a deep learning-based text understanding language which is aimed to help Facebook in understanding text in a near-human accuracy. The AI learns contextual content by scanning several thousands of posts per second with the ability to understand more than 20 languages. According to Facebook, DeepText leverages several deep neural network architecture that include "convolutional and recurrent neural nets", and can perform "word-level and character-level based learning."
The AI is built on, and extended the ideas of deep learning which were originally developed in papers by Ronan Collobert and Yann LeCun from Facebook AI Research.
Facebook uses its FbLearner Flow and Torch for DeepText's model training. Because the trained models are served with a click of a button through the FBLearner Predictor platform which has a distribution infrastructure model, Facebook's engineers can build new DeepText models through the self-serve architecture that DeepText provides.
Some examples that Facebook could do with DeepText, is to make it better in suggesting posts to users, and to also able to optimize Facebook's different user-facing products. Facebook can also use it to stop spam by creating more powerful filters that are better in differentiating good posts and bad posts.
But in order to do that, just like what other AIs should always do, DeepText needs to understand what it has to understand. And for that, it needs to learn by feeding on data.
While the start could be rough, as the AI learns more, the deep learning could reduce Facebook's reliance on language-dependent knowledge because the system can learn from text inputs with little to no preprocessing. This will help Facebook in scaling up quickly as it help the system in spanning multiple languages quickly with minimal engineering efforts.
AI's basic level of learning words is based on hard facts. With it, it needs to learn to understand the difference between meanings and potentially ambiguous words and phrases. What's more, it needs to know slang phrases and words that humans usually use to communicate, and should be able to solve tricky scaling and language challenges.
Learning From Users As They Use Facebook
Facebook is already able to recognize the faces of its users - Using DeepFace for example, it can see what we see. And because Facebook is considered as one of the biggest library of photos, it's AI has all the visual data it needs to learn.
But images are just one thing that makes Facebook. Text is another.
While a picture may worth a thousand words and videos are worth thousands of pictures, text means a lot in a different meaning and perspective. Because the ratio of text is certainly high on Facebook, the company wants to automate its system to make it as if it is one of the crowd.
At some point, AI is yet to be at that level of human in text understanding. But with DeepText operating in more than 20 languages, the team at Facebook said that it's already using Messenger to help DeepText learn the meaning of messages. What this means, DeepText, just like DeepFace, has all the data it needs to learn.
And what this further means, it shows that Facebook has used its popular messaging app as a way for quietly feed DeepText with users' messages.
In Messenger, DeepText is used by the AML Conversation Understanding team to get a better understanding of when someone might want to go somewhere.
"It's used for intent detection, which helps realize that a person is not looking for a taxi when he or she says something like, 'I just came out of the taxi,' as opposed to 'I need a ride'," wrote Facebook.
Because humans use many ambiguous words and abbreviations, Facebook relies its DeepText with deep learning so it can use "word embeddings" instead of the traditional NLP approaches where words are converted into a format that computer algorithms can learn. While NPL requires each word to be seen with exact spellings in the training data to be understood, deep learning's word learn by semantics by grouping similar words in a space.
Word embeddings is a mathematical concept that preserves the semantic relationship among words. So for example, Facebook can see that the word embeddings of "brother" and "bro" are close in space. According to Facebook, this is possible when the inputs are calculated properly.
This type of representation allows the company to capture the deeper meaning of words.
And because word embeddings can help Facebook in understanding the same semantics across different languages, Facebook can map words and phrases into one common embedding space. Here, DeepText should be able to build models that are language-agnostic.
Better Understanding Of Users, Better Representation Of Products And Ads
The ultimate goal for DeepText AI is to better understand what contents people share to then know what people actually want to see.
When it succeed in learning what it has to learn, DeepText has the potential to further improve Facebook experiences, not to mention its ability to understand users by a margin further. And not limited to that, DeepText in better understanding posts will enable Facebook to extract intent, sentiment and entities using a mixed content signals from both text and images.
This could be used to also improve its ad-targeting capability.
As an ad-supported and funded titan like Facebook, understanding what users want can have a huge benefit to the company's overall business performance.
DeepText is a project on progress. Facebook is continuing its effort to advance the AI technology and its application in collaboration with the Facebook AI Research Group.
While the potential of AI for Facebook is certainly big, users may see this as something that is utterly disconcerting. But again, Facebook lives on data and thrives as more people use its service. So the more the people are using Facebook, DeepText should become more intelligent. And by having an intelligent AI that learns by 'quietly listening' to users, that is a risk Facebook users should take when using the service.
Further reading: Google And Facebook In Creating The Future Of AI: Changing The Overall Technology Trends