rungjoo / plug-and-play-prompt-persona

P5: Plug-and-Play Persona Prompting for Personalized Response Selection (EMNLP 2023)

P5: Plug-and-Play Persona Prompting for Personalized Response Selection

The overall flow of P5 model.

Requirements

1. Pytorch 1.8
2. Python 3.6
3. transformers>=4.4.0
4. datasets
5. scipy

Dataset

• download link
• PERSONA-CHAT: positive 1, negative 19 (given)
• Focus: positive 1 (given), context negative sampling 2, random negative sampling 17

Dataset folder location

• ./dataset/FoCus
• ./dataset/personachat

Pretrained Model

• download_link

Model folder location

• ./model/NP_focus/roberta-base/model.bin
• ./model/NP_persona/roberta-base/model.bin
• ./model/NP_persona/roberta-large/model.bin
• ./model/prompt_finetuning/roberta-{size}_{personachat_type}/model.bin

NP_focus (NP: No Persona)

standard response selection model for Focus

cd NP_focus
python3 train.py --model_type roberta-{size} --epoch 10

• size: base or large

NP_persona

standard response selection model for PERSONA-CHAT

cd NP_persona
python3 train.py --model_type roberta-{size} --persona_type {persona_type} --epoch 10

• size: base or large
• persona_type: original or revised

SoP (Similarity of Persona)

Zero-shot baseline

• test_focus.py: testing for focus dataset
• test_perchat.py: testing for personachat dataset

cd SoP
python3 test_perchat.py --model_type roberta-{size} --persona_type {persona_type} --persona simcse --weight {weight} --agg max

• size: base or large
• persona_type: original or revised (It doesn't matter in focus))
• weight
  • 0.5 for original persona and Focus
  • 0.05 for revised persona

prompt_finetuning

Fine-tuned P5 model

• train.py: The main training file mentioned in the paper
• train_no_ground: don't use persona grounding
• train_no_question: don't use prompt question
• test_zeroshot.py: no persona during test

cd prompt_finetuning
python3 train.py --model_type roberta-{size} --data_type {data_type} --persona_type {persona_type} --persona {persona} --num_of_persona {num_of_persona} --reverse

• size: base or large
• data_type: personachat or focus
• persona_type: original or revised (It doesn't matter in focus)
• persona: simcse or nli or bertscore (recommend: simcse)
• num_of_persona: 1 to 5 (recommend: 2)
• reverse: option (order of persona sentences)

prompt_persona_context (ablation study)

Zero-shot P5 model

• test.py: The main test file mentioned in the paper
• test_no_ground.py: don't use persona grounding
• test_no_question.py: don't use prompt question
• test_other_question.py: variant of prompt question
• test_random_question.py: random prompt question
• dd means: When SRS is trained with dailydailog

cd prompt_persona_context
python3 test.py --model_type roberta-{size} --data_type {data_type} --persona_type {persona_type} --persona {persona} --num_of_persona {num_of_persona} --reverse

• size: base or large
• data_type: personachat or focus
• persona_type: original or revised (It doesn't matter in focus)
• persona: simcse or nli or bertscore (recommend: simcse)
• num_of_persona: 1 to 5 (recommend: 2)
• reverse: option (order of persona sentences)

Citation

@article{lee2023p5,
  title={P5: Plug-and-Play Persona Prompting for Personalized Response Selection},
  author={Lee, Joosung and Oh, Minsik and Lee, Donghun},
  journal={arXiv preprint arXiv:2310.06390},
  year={2023}
}