luheng/deep_srl

Implements an end-to-end SRL system using deep BiLSTM networks with variational dropout, supporting both gold and predicted predicates across CoNLL-2005 and CoNLL-2012 datasets. Built on Theano with GloVe embeddings, it provides pre-trained single models and 5-model ensembles for improved accuracy. The framework handles predicate identification and semantic role tagging as separate trainable components, enabling flexible deployment from interactive prediction to batch processing on CPU or GPU.

334 stars. No commits in the last 6 months.