DeText is LinkedIn’s open-source deep neural text understanding framework for ranking and classification tasks. Released under the BSD-2-Clause license, it has gathered about 1,300 stars on GitHub and reflects the technology behind large-scale search and recommendation features at LinkedIn.
DeText provides a modular framework for building text understanding models that power ranking and classification in production systems. It breaks the problem into reusable components: text encoding, interaction between queries and documents, and task-specific heads for scoring or labeling. Multiple encoder architectures are supported, from CNN and LSTM networks to BERT-style transformers, so teams can trade accuracy against latency for their use case. Because it was designed for industrial search and feed ranking, it emphasizes clean configuration, reproducible training pipelines and features that matter at scale, such as text fields, user context and in-batch negatives.
DeText organizes a model as representation, interaction and scoring stages, which makes it straightforward to experiment with different text encoders while keeping the rest of the pipeline fixed. It supports pointwise learning objectives for both ranking and classification, and LinkedIn open-sourced it to share the architecture family behind features such as search ranking and content classification, where careful text matching between a short query and long documents is the core problem.
DeText is free and open source under the BSD-2-Clause license. There are no paid plans; you run the framework on your own infrastructure and train models with your own data.
DeText is for search and NLP engineers who need a battle-tested framework for ranking, text matching or classification, especially teams working in TensorFlow-based environments similar to LinkedIn’s stack.