The social tech giant Facebook is open-sourcing its AI bot-building research. Using its Facebook AI Research (FAIR) lab's mission, the company plans to provide better access to researchers and engineers so they can build better bots.
The code library Facebook is open-sourcing is called "fastText", a library of code for text representation and classification that it uses to power its bots. According to Facebook, fastText that has been released on GitHub for people to download, only requires a compiler with "good C++11 support."
Facebook said that the new open-source technology can be "trained on more than 1 billion words in less than 10 minutes using a standard multicore CPU. Facebook's fastText can also classify a half-million sentences among more than 300,000 categories in less than five minutes."
Here, Facebook highlights speed and efficiency. These two are indeed the main appeal of fastText. With it, people can train bots a lot faster than using other methods.
In a statement, FAIR said:
The internet is a humble world full of words. Before there were search engines and directories to categorize them, the internet was a dark place. When Google came with a way to crawl the web, it has capitalized on this abundance of words with its algorithms.
Because the source of information is almost limitless, Google can create its own filtering to organize and rank websites and web pages, based on users' queries. This again, further leverage its linguistic capability.
Google is undisputed here, but Facebook is trailing not far behind.
Facebook is a social network giant that lives inside its own territory. It's walled-garden has become a paradise for information that are contained and managed by Facebook itself. Needless to way, Facebook has an internal search engine and spam filtration system that become more impressive as the company progresses.
For more than a while, deep neural networks have created a chance for humans to embrace the future of AI. But these systems, as effective as they can be, are still inefficient. They can take hours or even days to train. And in most cases, neural networks still require manual and supervised training. This is time-consuming, and aren't great for close-future projects.
On Facebook, there are billions of contents shared every day. To keep up with the fast growing data, Facebook has been using a variety of tools in its disposal to classify those contents. One of the tools the company uses is its deep neural network. While it is certainly accurate and useful, but the method requires a lot of training and a lot of time.
Another method Facebook uses in its effort to classify accurately and easily, is using fastText. Developed by FAIR, the artificial intelligence (AI) tool supports both text classification and learning word vector representations via techniques like bag of words and sub-word information.
Its main focus is by classifying words and sentences in order to create libraries that other programs can use as a reference when they're executing tasks.
"In order to be efficient on datasets with a very large number of categories, fastText uses a hierarchical classifier, in which the different categories are organized in a tree, instead of a flat structure (think binary tree instead of list)," said Facebook authors Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov in a post.
The bag of words process is faster because it essentially ignores word order. Instead, it focuses on the occurrences of words that represented in a multidimensional space. Linear algebra is then used to calculate the relations between a given query and the categorized set of words.
Using this technique, fastText is faster than deep learning methods, and can be incredibly good at understanding huge blocks of text. If compared to to deep learning models, fastText could cut training time from days to just seconds.
And what makes fastText handy is also because it can work with other languages, including: German, French, Spanish and Czech.
There were about 18,000 bots on Messenger in July 2016. By open-sourcing fastText, Facebook expects the number to grow significantly.