Hard on the heals of declaring its PyTorch tensor and neural network library “production ready” Facebook has open sourced the PyText natural-language processing modelling framework.
PyText has been released on GitHub under a BSD license, with pretrained models and tutorials Facebook said would help devs and datascientists build and train PyText models at scale.
There are models for common NLP tasks such as text classification, word tagging and semantic parsing, while contextual models tackle conversational understanding in NLP tasks. Facebook has also built two contextual models using PyText: a SeqNN model for intent labelling and a Contextual Intent Slot mode.
The social giant, which has a history of releasing code to the open-source community, reckoned it’s released PyText to make it easier to build and deploy NLP systems.
According to Facebook here, it “Blurs the boundaries between experimentation and large-scale deployment.”
Facebook claimed it’s used PyText to “iterate quickly on incremental improvements” and improve accuracy in NLP models in Portal – Feacebook’s new video calling device.
It claims to have improved accuracy of models in core domains by between five and 10 per cent with training times reduced by between three and five times.
PyText uses PyTorch and connects to ONNX and Caffe2, letting you convert PyTouch models to ONNX to be then exported to Caffe2. According to Facebook, PyText provides a “flexible, modular workflow with configurable layers and extensible interfaces.”
Facebook claimed exporting to Caffe2 provides the “performant and efficient” multithreaded C++ back end to serve “huge volumes of traffic efficiently with high throughput.”
NLP on mobile is an area of continued research for Facebook. The company’s plans include supporting multilingual modelling and other modelling capabilities, making it easier to debug models and – basically – optomising the software.