hkust-knowcomp/cat

Code for the ACL2023 paper: CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning (https://aclanthology.org/2023.acl-long.733.pdf).

CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning

This is the official code and data repository for the ACL2023 (Main Conference) paper: CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning.

1. Download Dataset/Model Checkpoints

The AbstractATOMIC dataset, including both annotated part and CAT's pseudo-labeled part, and the dataset for training event conceptualization generators are available at this link.

Our finetuned DeBERTa-v3-Large and GPT2 model checkpoints for four tasks are available at this link.

2. Required Packages

Required packages are listed in requirements.txt. Install them by running:

pip install -r requirements.txt

3. Training Different Components of CAT

To train the discriminative models of CAT, all training and evaluation codes are provided in event_concept_dicrimination and triple_discrimination folders. The training code are named as train.py in each folder and the parameters are all set as default.

python train.py --ptlm [microsoft/deberta-v3-large or our checkpoint weight]

Change the ptlm argument to select different pretrained language models or load our fine-tuned checkpoints.

To train the event conceptualization generator and COMET, use the codes provided in generator folder.

To train the event conceptualization generator, use the following command:

GPU=0 \
IN_LEN=40 OUT_LEN=46 OUT_DIR=./GPT2-Conceptualization-Generator \
SAVE_EVERY_EPOCH=False DO_PRED_EVERY_EPOCH=False EVAL_EVERY=100 \
USE_NL_RELATION=True TRAIN_BATCH_SIZE=32 VALID_BATCH_SIZE=32 \
DO_TRAIN=True DO_PRED=False TRAIN_EPOCHS=50 TOKENIZER=gpt2-xl GPT2_MODEL=gpt2-xl \
TRAIN_DATA_PATH=[Training Data Path] \
DEV_DATA_PATH=[Validation Data Path] \
TEST_DATA_PATH=[Test Data Path] \
PRED_FILE=[Prediction File Path] \
REPLACE_BRACKET=True \
python models/comet_atomic2020_gpt2/comet_concept_gpt2.py

The training, validation, and testing data can be downloaded at this link.

To train COMET, use the following command:

GPU=0 \
IN_LEN=24 OUT_LEN=40 OUT_DIR=./COMET \
SAVE_EVERY_EPOCH=False DO_PRED_EVERY_EPOCH=False \
EVAL_EVERY=100 USE_NL_RELATION=True TRAIN_BATCH_SIZE=32 VALID_BATCH_SIZE=32 \
DO_TRAIN=True DO_PRED=False TRAIN_EPOCHS=50 TOKENIZER=gpt2-xl GPT2_MODEL=gpt2-xl \
TRAIN_DATA_PATH=[Training Data Path] \
DEV_DATA_PATH=[Validation Data Path] \
TEST_DATA_PATH=[Test Data Path] \
PRED_FILE=[Prediction File Path] \
REPLACE_BRACKET=False \
python models/comet_atomic2020_gpt2/comet_concept_gpt2.py

The data format follows the same format as the original input of COMET, details can be found at here.

4. Citing this work

Please use the bibtex below for citing our paper:

@inproceedings{CAT,
  author       = {Weiqi Wang and
                  Tianqing Fang and
                  Baixuan Xu and
                  Chun Yi Louis Bo and
                  Yangqiu Song and 
                  Lei Chen},
  title        = {CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning},
  year         = {2023},
  booktitle    = {Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, {ACL} 2023}
}

5. Acknowledgement

The authors of this paper are supported by the NSFC Fund (U20B2053) from the NSFC of China, the RIF (R6020-19 and R6021-20), and the GRF (16211520 and 16205322) from RGC of Hong Kong, the MHKJFS (MHP/001/19) from ITC of Hong Kong and the National Key R&D Program of China (2019YFE0198200) with special thanks to HKMAAC and CUSBLT. We also thank the UGC Research Matching Grants (RMGS20EG01-D, RMGS20CR11, RMGS20CR12, RMGS20EG19, RMGS20EG21, RMGS23CR05, RMGS23EG08).