detext

Why we included this project

DeText comes from LinkedIn's search team and gives you a framework for training deep neural text models, with ranking and classification in search and recommender systems as the main targets. Rather than a fixed model, it lets you assemble the network from parts: word embeddings, text encoders like CNN, LSTM, or BERT, an interaction layer that computes cosine similarity or concatenation between two texts, and a wide-and-deep head. Configuration happens through command-line parameters, so you can iterate on architectures and hyperparameters without hand-wiring TensorFlow graphs each time. If you work on query-to-document relevance, query intent classification, or auto-completion, it is directly applicable, since it includes learning-to-rank losses, tf-ranking integration, and tutorials on real public datasets. The main caveat is that it targets model training with TensorFlow 2.4 and Python 3.7, so it suits teams that run their own training pipelines rather than looking for a hosted service.