ege-RoBERTa

Learning event graph knowledge for abductive reasoning.

To this end, we involve a two stage learning process to introduce the event graph knowledge, and a variational autoencoder based model ege-RoBERTa to capture the event graph knowledge.

Pre-training Stage:

Learning Event Graph Knowledge from a Pseudo Instance Set

• Preprocess datasets using pret.py
• Using event_order_gen.py to sample adjacent and non-adjacent event pairs for traing a next event prediction model. Then run learning_event_order.sh to train the next event prediction model (described in the Sec 5.2 of original paper).
• Using aux_dat_gen.py to construct the pseudo instance set.
• Then pretrain.sh is used for conducting the first stage training process.

Finetuning Stage:

Adapt Event Graph Knowledge to the Abductive Reasoning Task

Please refer to train_anli.sh and Train_anli.py.

Model Architecture

Files for constructing model architecture is contained in the file folder onmt.