Facebook turned a proprietary natural language processing (NLP) code out for public consumption last week, enlisting third-party developers to push the social-media platform into the market for consumer communications.
Facebook’s move to open the source code for the PyText NLP it uses to process billions of voice-scripted messages each day comes as the the platform operator is wading into the manufacture and sale of devices that carry them.
In October, the company launched a line of devices designed to optimize the experience of communicating via its Messenger instant-messaging service. Called Portal, the smart screens track users as they move in space, and let them access wireless services and digital assistants in addition to their Facebook accounts.
In the desktop and bookshelf units, Facebook's PyText NLP technology recognizes users and improves accuracy when executing their voiced commands. The code library also is useful in research and development for NJP applications, with the PyText libraries housing templates for building, modeling and testing that Facebook is keen to augment as voice-to-machine use-cases grow.
PyText is forked from PyTorch, a code for machine learning that Facebook built using language elements of Python for programming and Torch for scripting in low-latency applications.
The California-based company’s AI research group developed PyTorch, which uses GPUs and tensor cores to accelerate deep learning, before they turned out the source code to the development community two years ago. The group also developed a PyText predecessor called DeepText, which possesses limited functionality for statistics-driven ML.
Customizable Machine Learning
Applications that derive meaning from human speech consume enormous amounts of computing in development, and rely on large amounts of memory when deployed. According to Facebook, PyText lets developers optimize resources across each cycle, with its machine-learning tech helping developers achieve a virtuous circle of faster iteration and implementation.
Released to the GitHub platform, PyText libraries contain tools and templates for semantics parsing, sequence tagging and multi-task modeling. The code improves predictive capabilities in statistics-based ML, with the better pattern recognition realized in distributed GPU computes working to lower training times.
Facebook says that PyText allows developers to combine out-of-the-box components, which can then be adjusted and augmented to suit the given application. The modular approach permits precision control for faster testing of models built with ML.
Compatibility with ONNX, the open neural-network exchange, lets developers use PyText in different AI frameworks. When exporting PyText code to the Caffe2 framework, developers can change components at speed and scale their applications for faster deployment.
The company says developers can use the code to build neural networks with dynamic structures for conversational artificial intelligence. Facebook points to in-house use of PyText applications that daily make billions of low-latency predictions in dozens of human languages as proof of the code’s robust scalability.
Applications Continue Downstream
Applications written with PyText are at work in the Portal device line. The two models Facebook brought to market in early November use AI and ML in a number of ways, including for device security and data protection.
‘Hey, Portal’ activation and command functionalities rely on NLP to parse instructions and direct calls out from the devices, which can connect via Messenger IM to phones, computers and other Portal devices. Scaling the ability to parse, classify and execute to production-level with PyText NLP tech enabled the company to bring the devices to market, Facebook says.
Portal users direct the device’s onboard smart camera with NLP, and they can call up augmented-reality functions and filters that enhance the user experience. They can access internet-based music services with NLP and communicate with Amazon’s Alexa digital assistant via a dedicated interface.
Enhanced NLP functionality is limited mainly to Portal-to-Portal communication, meaning that the devices won't soon precipitate Facebook's move from an advertising-dominated revenue model. At least, not without significant buy-in from the platform's 1.3 billion regular users.
In order to drive uptake, including of PyText, Facebook says it's working on end-to-end workflows for mobile devices and is looking to the third-party developers to expand models for multilingual applications.