This work sets out to acquire Hebrew exemplar sentences with FrameNet annotations by projecting annotations from English. To this end, we use the OpenSubtitles 2016 dataset of aligned English-Hebrew subtitles of movies and television shows.

This repository contains the following:

• Thesis in PDF format
• Source code for the SRL visualization tool
• Jupyter notebooks for dataset construction, dataset statistics, and feature extraction
• Two sqlite databases: manual_annotations.sqlite3 (original seed) and data.sqlite3 (result of classifier annotation on a subset of the dataset)
• The code for the classifier, and a pickled trained classifier

In order to run the SRL visualization tool, you need:

• Python 3.5 or later (Python 2 not supported)
• conllu package to parse CoNLL-U format
• scikit-learn and imbalanced-learn for the classifier
• Download the dataset from here and place it in the static/dataset directory

The best way to get started is to create a new virtual environment (either through Conda or python3 -m venv <venv name>), and run pip install -r requirements.txt, followed by python3 viz.py.

The following diagram shows the data pipeline in our work:

The computer used to process the data is an Intel Xeon E5645 @ 2.40 GHz with 24 cores and 128GB RAM.

• English POS-tagging and dependency parsing: less than one day
• Hebrew segmentation, morphological analysis, morphological disambiguation, and dependency parsing using YAP: one month
• Automatic English-Hebrew alignment: one day
• English SRL: one week
• Each of the other parts took a few minutes