"DeepText" to understand sentences like human beings, "from taxi arrangements" to "content filtering"

ByKate Ter Haar

There are various contents such as text, photos, movies, etc. on Facebook, but the most number is definitely text. Understanding such numerous existing texts, Facebook's developers have been conducting research as "Can you make something that can help the user?" Then, it is possible to display the content desired by the user in the place where the user can see, filter unnecessary contents,Deep learningBased text understanding engine "DeepTextThe development of "is progressing.

Introducing DeepText: Facebook's text understanding engine | Engineering Blog | Facebook Code | Facebook
https://code.facebook.com/posts/181565595577955/introducing-deeptext-facebook-s-text-understanding-engine/


Facebook's new DeepText AI categorizes everything you write | TechCrunch
http://techcrunch.com/2016/06/01/facebook-deep-text/

"DeepText", a text understanding engine developed by Facebook, is based on a multilayered neural network architecture and can be called a kind of artificial intelligence (AI) that can also learn at the word level or at the character level. It is possible to understand multiple elements in the text in a form close to the actual human being, and it seems that it is possible to analyze thousands of text contributions per second. DeepText can understand what the user wants from the text, for example, if the user writes "There is no way to return from the station of ○ ○", so that the user responds "Do you want to call a taxi?" He said he will be able to propose the intention included in the text.

DeepText already supports more than 20 languages ​​at the time of article creation, but since the Facebook community is very global, it is necessary to analyze many languages ​​in order for DeepText to be more "usable" service. In addition, it is necessary to learn slang and spelling mistake etc. so that it can correspond to various text forms, but by using deep learning to reduce the trouble of learning language dependent knowledge as much as possible He said that he was successful.

DeepText has already started testing on Facebook. For example, if you incorporate DeepText into "Messenger", Facebook's messaging application, you can analyze the text entered by the user and recommend a taxi to the user if it deems it necessary.

The way the message application actually reacts to the text entered by the user, suggesting or not recommending a taxi can be seen in the following movie.

While exchanging messages with friends via Facebook Messenger application


When you enter "I need a ride", the button "Request a Ride" appears immediately below it. Tap this to instantly arrange a taxi. On the other hand, there is no response to the text "I do not need a ride" just below.


Even if you input "Take a taxi" or "Let's get on the car from here", a taxi arrangement button appears. If you look closely well, it seems like you are arranging a taxi with the words "taxi" "cab" "ride", but "I like to ride donkeys (I like to ride a donkey)" If you enter, you can see that the ordering button of the taxi is not displayed and its accuracy is quite good.


In addition to its application in the messenger application, Facebook also uses it as a tool to find the user's purpose. For example, suppose a Facebook user posts "I want to sell my old bike for 200 dollars, who is interested?" When DeepText analyzes this post, it understands "I'm trying to sell something", extracts only meaningful information (something that you are trying to sell or price) from the submitted posts, and the seller sells it You can introduce existing tools that will make it easier. In addition, it is possible to analyze the user's contribution by "people", "place", "event", etc., and to distribute contents suitable for each user.

In addition, many celebrities on the Facebook communicate with the general users, but when they communicate, thousands of comments from around the world gather, so excellent comments will tend to be buried. However, if you use DeepText, you will be able to find excellent ones from a number of comments and display it preferentially.

It is said that development of DeepText is being carried forward in collaboration with AI research group of Facebook, and the project will be "to better understand what the user is interested in" "Understand the text and connect it with visual content "And" to build a new deep layer neural network architecture ".