Semantic Role Labeling demo
This is a demonstration of various state-of-the-art1 semantic role labeling models that I developed based on the excellent mate-tools NLP pipeline and other tools2.
Any problems or questions? Please contact Michael Roth. Downloads can be found here
1For English (PropBank/NomBank), the model achieves a F1-score of 87.9 percentage points, the best result reported on any SRL data set to date.
For English (FrameNet), the model achieves a F1-score of 79.3% (given gold frames), an improvement of 2.4 points over previous work.
The German model achieves a F1-score of 81.4%, the current state-of-the-art result and the first to crack the 80.0 point mark.
Stanford CoreNLP (tokenization, sentence splitting, coreference; Manning et al., ACL'14),
mate-tools (preprocessing, predicate/target identification; Bohnet, COLING'10; Björkelund et al., COLING'10),
More information can be found in the GitHub code repositories here and here (FrameNet) as well as in the PathLSTM and framat(e) papers here.