TemporalBART

The source code for the paper "Conditional Generation of Temporally-ordered Event Sequences".

Requirements

• torch 1.6.0
• allennlp 1.1.0
• transformers 3.0.2
• rouge-score
• graphviz
• lemminflect
• pandas

Data

In this work, we specifically collect our training data from the EventsNarratives corpus, and evaluate our models in a zero-shot manner on CaTeRS and MCTaco dataset.

• The preprocessed training data can be downloaded from here, where the 100K version is used for our best models.

• The preprocessed evaluation data can be found here: CaTeRS, MCTaco

Training

To train the TemporalBART models described in the paper, first change the data paths (train_data_path, validation_data_path, test_data_path) in the config file to where you store the downloaded data, then run the following commands:

• TemporalBART: allennlp train ./configs/temporal_bart_config.json -s /PATH/TO/MODEL_CHKPT/ --include-package denoising_event_lm -f

• TemporalBART-indexed: allennlp train ./configs/temporal_bart_indexed_config.json -s /PATH/TO/MODEL_CHKPT/ --include-package denoising_event_lm -f

Evaluation

To evaluate the fine-tined models on temporal event ordering:

• CaTeRS:

allennlp evaluate /PATH/TO/MODEL_CHKPT_TARGZ caters_entity_chains_seqdata_scrambled2.pkl --cuda-device 0 --include-package denoising_event_lm -o '{"validation_dataset_reader": {"chain_len_min": 2, "event_del_prob": 0.0, "event_del_all_prob": 0.0, "allow_empty_events": false}}'

• MCTaco:

python3 denoising_event_lm/predictors/event_lm/test_demo_event_lm_mctaco_before_after.py \
    --archive-path /PATH/TO/MODEL_CHKPT_TARGZ \
    --predictor demo_denoising_event_lm_mctaco_before_after_predictor \
    --include-package denoising_event_lm \
    --cuda-device 0 \
    -o '{}' --input-path mctaco_event_ordering_before_after.json --beams 1 --feed-unseen > output.txt

To evaluate the fine-tuned models on ordering unseen events:

python3 denoising_event_lm/predictors/event_lm/test_demo_event_lm_orderextra.py \
    --archive-path /PATH/TO/MODEL_CHKPT_TARGZ \
    --predictor demo_denoising_event_lm_orderextra_predictor \
    --include-package denoising_event_lm \
    --cuda-device 0 \
    -o '{}' --input-path caters_entity_chains_seqdata_scrambled2.pkl --beams 2 --chain_len_min 2 > output.txt

Model Checkpoints

The final models used in our paper can be downloaded with the following links:

• TemporalBART
• TemporalBART-indexed