Code associated with:

Temporal Information Extraction by Predicting Relative Time-lines

Artuur Leeuwenberg & Marie-Francine Moens In Proceedings of EMNLP, Brussels, Belgium, 2018.

Terms of Usage

Apache License 2.0

Please cite the paper when using the code.

What this code can be used for:

Training relative time-line prediction models (TLM), trained from TimeML-annotated data, and to convert TimeML-style annotations to relative time-lines (TL2RTL), i.e. to reproduce the results from the EMNLP 2018 paper. The data used in the original paper was added to facilitate reproducing the results from the paper.

How do I get set up?

1. Install Python 3.5 (if not yet installed)
2. Setup and activate (line 2) a virtual environment, and install the Python dependencies with:

virtualenv venv -p python3.5
source venv/bin/activate
pip install -r requirements.txt

3. Unzip the word vectors in data/wordvectos/xaa to obtain data/wordvectors/glove.6B.50d.w2vf.txt

Reproducing the results from the paper:

To produce the results from Table 4 from the paper, you need to run two scripts:

sh EXP-TLM.sh
sh EXP-TL2RTL.sh

The first will rerun the S-TLM and C-TLM experiments in the different loss settings (evaluation scores can be found at the end of the corresponding log files). And the second will rerun the TL2RTL experiments.

!!! By default these scripts run many subprocesses simultaneously. If you prefer to run the experiments one-by-one you should remove the &-symbol at the end of the python calls in the scripts. You can indicate which GPU you want to use by setting the CUDA variable in the scripts (-1 for using CPU instead).

Obtaining more information

To get more information on how to use the scripts you can inspect the .sh files or run:

python main.py -h

for TLM options or

python timeml_to_timeline.py -h

for TL2RTL options.

Questions?

Any questions? Feel free to send me an email!