Social Networking giant Facebook has been running on its own Natural Language Processing (NLP) framework, PyText, for conquering quick testing and large-scale deployment challenges. PyText is a deep-learning based NLP modeling framework built on PyTorch. In an attempt to assist developers develop and practice NLP systems, the social networking firm has decided to also open source the PyText framework, with serving its pre-trained models and tutorials.
With PyText Facebook has been able to attain rapid experimentation, tackle text processing and vocabulary management at scale, and strap the PyTorch ecosystem for prebuilt models and tools, as per the company statement. In addition, the company noted that they have utilized this framework to take NLP models from idea to full execution in just days, in place of weeks or months, and to set up compound models that depend on multitask learning. PyText, at Facebook today is utilized for over a billion daily predictions, exhibiting that it can operate at production scale and still meet stringent latency needs. Facebook described that typically, researchers and engineers have to tradeoff between frameworks developed for tests and frameworks built for production. This is mainly true for NLP systems that can require developing, training, and experiment dozens of models, and which use an intrinsically dynamic structure.
With the power of PyTorch 1.0 that addressed research and production hurdle with a single unified framework, PyText is able to convey PyTorch’s 1.0 features into NPL. That features include the ability to distribute models across various businesses within the AI community, pre-built models of common NLP tasks like text classification and language modeling, and contextual models to enhance conversational understanding. The social networking giant plans to utilize the framework in its own solutions to give significant features, flag policy-violating posts, perform translations and improve products. Facebook, looking ahead, plans to address end-to-end workflows for on-device models and offer multilingual modeling along with other modeling capabilities, which give the ability to fix and advance distributed training.