stevenlol / pathlstm

Neural SRL model

PathLSTM

This repository contains code for the PathLSTM semantic role labeler introduced in Roth and Lapata, 2016. It is built on top of the mate-tools semantic role labeler. The PathLSTM model achieves state-of-the-art results on the in-domain (87.9) and out-of-domain (76.6) test sets of the CoNLL-2009 data set.

Dependencies

The following libraries and model files need to be downloaded in order to run PathLSTM on English text:

• Bernd Bohnet's dependency parser and model files (anna-3.3.jar and CoNLL2009-ST-English*.model)¹
• The WSJ tokenizer from Stanford CoreNLP (stanford-corenlp-3.x.jar)
• The most recent PathLSTM SRL model (July 2016), available from Google Drive here

Running PathLSTM

If copies of all required libraries and models are available in the subdirectories lib/ and models/, respectively, PathLSTM can simply be executed as a standalone application using the script scripts/parse.sh. These scripts run necessary preprocessing tools on a given input text file (assuming one sentence per line), and apply our state-of-the-art model for identifying and role labeling of semantic predicate-argument structures.

It is also possible to apply the PathLSTM model on already preprocessed text in the CoNLL 2009 format, using the Java class se.lth.cs.srl.Parse. Since PathLSTM is trained based on preprocessed input from specific pipelines, however, we strongly recommend to use the complete pipeline to achieve best performance.

References

If you are using PathLSTM in your work--and we highly recommend you do!--please cite the following publication:

Michael Roth and Mirella Lapata (2016). Neural Semantic Role Labelling with Dependency Path Embeddings. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics. To appear.

If you are using the built-in preprocessing pipeline, please also cite the following publication:

Bernd Bohnet (2010). Very high accuracy and fast dependency parsing is not a contradiction. The 23rd International Conference on Computational Linguistics (COLING), Beijing, China.

---

¹ To reproduce our evaluation results on the CoNLL-2009 data set, preprocessing components must be retrained on the training split only, using 10-fold jackknifing.