xdomain-temprel

Code for training and evaluating a TempRel classifier on MATRES, UDS-T, and distantly labeled examples.

Before running the model, make sure that the desired data files have been downloaded as per:

• timebank/README.md: MATRES Dataset
• udst/README.md: UDS-T Dataset
• beforeafter/README.md: Distantly-labeled BeforeAfter Examples
• timex/README.md: Distantly-labeled Timex-anchored Examples

Training & Evaluating

To train and evaluate on MATRES:

python train.py --lm roberta --data matres --output_dir /PATH/TO/MODEL_CHKPTS/ --epoch 5 --batch 32

python eval.py --lm roberta --data matres_dev matres_test --model_dir /PATH/TO/MODEL_CHKPTS/

To train and evaluate on UDS-T:

python train.py --lm roberta --data udst --output_dir /PATH/TO/MODEL_CHKPTS/ --epoch 5 --batch 32

python eval.py --lm roberta --data udst_dev_maj_conf_nt udst_test_maj_conf_nt --model_dir /PATH/TO/MODEL_CHKPTS/

To train and evaluate on timex train/test split from Goyal and Durrett, 2019:

python train.py --lm roberta --data distant --output_dir /PATH/TO/MODEL_CHKPTS/ --epoch 5 --batch 32

python eval.py --lm roberta --data distant_test --model_dir /PATH/TO/MODEL_CHKPTS/

To train and evaluate on additional timex data, sampled more evenly from sources in the English Gigaword Fifth Edition, specify a dataset from timex/data/*pkl, e.g.:

python train.py --lm roberta --data timex/data/d1k.pkl --output_dir /PATH/TO/MODEL_CHKPTS/ --epoch 5 --batch 32

For multiple data sources, simply specify as a space-separated list, e.g.:

python train.py --lm roberta --data matres timex/data/d1k.pkl --output_dir /PATH/TO/MODEL_CHKPTS/ --epoch 5 --batch 32

Fine-Tuned Models

You can download RoBERTa models that have been fine tuned:

• on MATRES: link
• on 1k MATRES examples and 10k DistantTimex examples: link