Developers have a new tool to help mobile apps understand text, thanks to a Facebook open source project update on Tuesday. The social networking company’s AI research group released a new version of FastText, a programming library that’s designed to make it easier for developers to deploy text-focused machine learning systems.
Using a technique the researchers are calling FastText.zip, developers can compact a language recognition model so that it takes up two orders of magnitude less memory while maintaining much of the accuracy they would get out of a non-compacted model. It’s a move that allows those models to be deployed on less powerful devices like smartphones and Raspberry Pis, making them more useful for a broader variety of applications.
In addition, Facebook released a pair of tutorials designed to help developers get started using FastText. The team also released a set of almost 300 pre-trained language sets to simplify matters further.
The goal behind FastText is to make it easier for people with a light background in programming to do text classification, (the process of assigning a block of words into a set of categories) and text representation (the process of turning unstructured text into numbers for computation).
“That was the idea behind the library — to make it a very accessible library for any text-related machine learning problems,” Facebook Research Scientist Armand Joulin said.
What makes FastText special is that the team building it at Facebook is focused on taking existing techniques and making them more accessible to everyday developers, so that it’s easier for people without a PhD in data science to implement machine learning in their apps.
For example, FastText can be used to power features like hashtag autocompletion, so that users can more quickly insert relevant tags into social media posts. It can also help with sentiment analysis, so that applications can understand whether users are saying something positive or negative.
FastText was also specifically built to handle a wide variety of languages, according to Edouard Grave, a postdoctoral fellow at Facebook. In particular, he said that it could handle languages like German and French that might cause problems for other systems.