Facebook has created a new modeling framework for natural language processing (NLP) systems open source. It’s about PyText, which has to close the gap between experimenting with such systems and getting them running. That’s what Silicon Angle reports.
According to the developers of Facebook, PyText is useful, because building the neural networks you need for NLP systems is traditionally often a tricky part. There is a difference between the frameworks optimized for experimentation and those optimized for production.
Building an NLP system often requires extensive training and testing of dozens of AI models. Much of the training and testing takes place on frameworks that focus on research. According to Facebook, these are useful because they offer a simple interface that makes making the models go faster.
But these frameworks are not suitable for running NLP systems that are in production. The reason is that the models that use such systems suffer from increased latency and memory usage. There are already production frameworks in place that can solve these problems, but they would make it difficult to create dynamic representations of text. This is a prerequisite for any reliable NLP system, according to the company.
That is why Facebook decided to develop the open source PyText. This is a software library built on the open source PyTorch deep learning framework. PyText is designed to meet the specific requirements of NLP model making. It does this by offering a simpler workflow that makes it possible to experiment faster. In addition, it provides access to various pre-created architectures for models, as well as tools for text processing and vocabulary management.
PyText can also have access to the wider PyTorch ecosystem. It has its own tools and models for NLP systems. PyText makes it possible to experiment with NLP systems as well as to put them into production for tasks such as document classification and semantic parsing.This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.