Semantic Role Labeling with subword composition

The code for our ACL18 paper "Character-Level Models versus Morphology in Semantic Role Labeling".

We provide sample training/testing scripts for different subword units under example_scripts folder.

Scripts Overview

Train SRL models on CoNLL-09 SRL training sets and test/evaluate trained models on CoNLL-09 evaluation sets for all languages.

1. simple_UNIT.sh: Trains/tests base SRL models for the given subword UNIT. Please check train.py for parameter descriptions.

2. ensemble_UNIT1_UNIT2_UNITn: Voting ensemble for the provided pretrained base SRL models (UNIT1, UNIT2, ..., UNITn).

3. sg_UNIT1_UNIT2_UNITn: Trains/tests a stack generalizer model from the predictions of pretrained base SRL models (UNIT1, UNIT2, ..., UNITn).

Testing Environment

Language : Python 2.7
CUDA : Cuda compilation tools, release 8.0, V8.0.44
Libraries: PyTorch 0.2.0 post 3, numpy 1.13.0