Code for COMMA (COLING 2022)

The Code of COMMA (COgnitive fraMework of huMan Activity).

Paper: COMMA: Modeling Relationship among Motivations, Emotions and Actions in Language-based Human Activities. Yuqiang Xie, Yue Hu, Wei Peng, Guanqun Bi, Luxi Xing. [pdf] in COLING 2022.

Setup

Download the pre-trained language models from https://huggingface.co/models.

• BERT-BASE/LARGE
• ROBERTA-BASE/LARGE
• GPT-2-LARGE
• BART-LARGE

Running the COMMA experiment

Please go into src file.

Emotion Understanding

# BERT
bash ./scripts/comma_m2e/run_comma_m2e_bert.sh
# RoBERTa
bash ./scripts/comma_m2e/run_comma_m2e_roberta.sh

Motivation Understanding

# BERT
bash ./scripts/comma_e2m/run_comma_e2m_bert.sh
# RoBERTa
bash ./scripts/comma_e2m/run_comma_e2m_roberta.sh

Conditioned Action Generation

GPT-2

# baseline
bash ./scripts/comma_x2b/run_gpt2_cag_baseline.sh
# train w/ emotions
bash ./scripts/comma_x2b/run_gpt2_cag_with_emotion_prediction.sh
# train w/o all
bash ./scripts/comma_x2b/run_gpt2_cag_wo_emotion_prediction.sh

BART-Large

Please return into BART file.

CUDA_VISIBLE_DEVICES=0 python finetune_trainer.py --model_name_or_path ./model/bart-large --data_dir ./bart_data/mber/ --output_dir ./output --learning_rate=3e-5 --do_train --do_eval --do_predict --evaluation_strategy steps --predict_with_generate --n_val 1000 --overwrite_output_dir --per_device_train_batch_size 8 --gradient_accumulation_steps 4

Acknowledgement

Thanks for the following github projects:

• https://github.com/huggingface/transformers
• https://github.com/allenai/allennlp
• https://github.com/debjitpaul/Multi-Hop-Knowledge-Paths-Human-Needs