Giters

Leezekun / dialogic

[EMNLP 2022] Code and data for "Controllable Dialogue Simulation with In-Context Learning"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Dialogic: Controllable Dialogue Simulation with In-Context Learning

This is the official pytorch implementation of our paper titled Controllable Dialogue Simulation with In-Context Learning.

Introduction

Starting from a small seed dataset, our method dialogic can generate good-quality annotated dialogues without human labor, parameter update, or engineering efforts, which is a much more time-saving and cost-efficient alternative to crowdsourcing in dataset creation.

We show a demo of how a dialogue is simulated above. You can type into your user goal or use the automatically generated one. We also provide simulated dialogues in the ./simulated_dialogues directory. The description of data format can be found here.

Table of Contents

Preparation

The code is placed in the ./code directory. As we use PPTOD as the auxiliary model for verification in this repo, most important files dialogic_*.py are in the ./code/pptod/E2E_TOD directory.

Environment setup

Set up the environment for PPTOD and SimpleTOD. To set up the environment for MinTL, please refer to ./code/MinTL/README.md.

pip install -r requirements.txt
python -m spacy download en_core_web_sm

Data preparation

We use MultiWOZ_2.3 dataset by default. MultiWOZ_2.0, MultiWOZ_2.1, and MultiWOZ_2.4 datasets are also supported. You can use the following script to prepare the data.

cd ./code/pptod/data/multiwoz
# MultiWOZ_2.3 dataset
chmod +x ./data_preparation23.sh 
./data_preparation23.sh 
# MultiWOZ_2.0 dataset
# chmod +x ./data_preparation.sh 
# ./data_preparation.sh 
# MultiWOZ_2.1 dataset
# chmod +x ./data_preparation21.sh 
# ./data_preparation21.sh 
# MultiWOZ_2.4 dataset
# chmod +x ./data_preparation24.sh 
# ./data_preparation24.sh

Auxiliary model preparation

We use PPTOD as the auxiliary model for verification in this codebase. To use it, you should download the initial checkpoint you want and unzip it in the ./code/pptod/checkpoints directory. We use PPTOD-small by default.

cd ./code/pptod/checkpoints
# Downloading Initial PPTOD-small Checkpoint:
chmod +x ./download_pptod_small.sh
./download_pptod_small.sh
# Downloading Initial PPTOD-base Checkpoint:
chmod +x ./download_pptod_base.sh
./download_pptod_base.sh
#Downloading Initial PPTOD-large Checkpoint:
chmod +x ./download_pptod_large.sh
./download_pptod_large.sh

Then you can use the script to train the verification model on the given seed dataset (1% few-shot setting by default):

cd ./code/pptod/E2E_TOD/sh_folder/small/training
chmod +x pptod_small_training_few_shot_0.01.sh
./pptod_small_training_few_shot_0.01.sh

Some important options include:

  • --train_data_ratio: the ratio of training data we use, i.e., the few-shot setting (1% by default).
  • --ckpt_save_path: the path where the trained verification model is saved.

We provide the checkpoints of verification models trained on 1%/5%/10% of the training data, which are placed at ./code/pptod/E2E_TOD/ckpt23/small/.

Simulation

Put your OpenAI API key in ./code/pptod/E2E_TOD/dialogic_utils.py to use GPT-3!

Dialogue simulation

First, we extract the turn-level annotations.

cd ./code/pptod/E2E_TOD/
# process the dialogues to get turn-level annotations
python dialogic_pre_process.py\
 --train_data_ratio 0.01

Then we generate the user goals, select in-context examples and construct the prompts.

cd ./code/pptod/E2E_TOD/
python dialogic_aug_e2e.py\
  --train_data_ratio 0.01\
  --augment_type combine\
  --augment_time 1\
  --k_shot 2\
  --temperature 0.2

Some important options include:

  • --train_data_ratio: the ratio of training data we use, the few-shot setting.
  • --augment_type: how to generate the user goals, options: [combine substitution, random].
  • --augment_time: how many times of the seed dataset we are going to augment.
  • --k_shot: how many in-context examples used in the prompt.
  • --temperature: the temperature when using the combine method. Lower temperature results in less random example selection.

Finally, you can use the following script to start simulating the dialogues:

cd ./code/pptod/E2E_TOD/sh_folder/small/simulation/
chmod +x ./pptod_small_few_shot_0.01_simulation.sh
./pptod_small_few_shot_0.01_simulation.sh

Some important options include:

  • --train_data_ratio: the ratio of training data we use, the few-shot setting.
  • --pretrained_path: the path of trained auxiliary verification model.
  • --output_save_path: the path to store the simulated dialogues.
  • --augment_dialog_path: the path where the constructed prompts store.
  • --max_turn_num: the maximum number of turns in each simulated dialog.
  • --max_dialog_num: the maximum number of dialogs simulated using gpt-3.
  • --max_aug_num: the size of simulated data w.r.t. the original seed data.
  • --verify_bs: whether to use the auxiliary verification model to correct the generated belief state annotations.
  • --verify_da: whether to use the auxiliary verification model to correct the generated dialog act annotations.
  • --debug: whether to print out the simulation process.
  • --save: whether to save the simulated dialogues.

You will see the dialogue simulation process as:

--------------------------------------------- Generation process of turn 0 ---------------------------------------------
GPT-3 generated user turn   >> You require([hotel] pricerange is cheap , day is tuesday , people is 5): hi there , i ' m looking for a hotel that is cheap and can accommodate 5 people on tuesday .
Verifier generated bs       >> [hotel] pricerange cheap and accommodate 5 day tuesday
Corrected bs                >> [hotel] day tuesday pricerange cheap people 5
Corrected user turn         >> You require([hotel] day is tuesday , pricerange is cheap , people is 5): hi there , i ' m looking for a hotel that is cheap and can accommodate 5 people on tuesday .
DB query result             >> [db_3]
Verifier generated da       >> [hotel] [inform] choice [request] area
Corrected system turn       >> Assistant([hotel] [inform] choice [request] area): i have [value_choice] options that are in the [value_area] .
------------------------------------------------ Conversation of turn 0 ------------------------------------------------
User                        >> hi there , i ' m looking for a hotel that is cheap and can accommodate 5 people on tuesday .
System                      >> i have [value_choice] options that are in the [value_area] .
------------------------------------------------------------------------------------------------------------------------

--------------------------------------------- Generation process of turn 1 ---------------------------------------------
GPT-3 generated user turn   >> You require([hotel] area is east): perfect , can you give me more information about the hotels in the east ?
Verifier generated bs       >> [hotel] pricerange cheap and accommodate 5 day tuesday
Corrected bs                >> [hotel] day tuesday pricerange cheap people 5 area east
Corrected user turn         >> You require([hotel] area is east): perfect , can you give me more information about the hotels in the east ?
DB query result             >> [db_2]
Verifier generated da       >> [hotel] [inform] area name internet parking
Corrected system turn       >> Assistant([hotel] [inform] area name internet parking): there is the [value_name] , which has free wifi and parking .
------------------------------------------------ Conversation of turn 1 ------------------------------------------------
User                        >> perfect , can you give me more information about the hotels in the east ?
System                      >> there is the [value_name] , which has free wifi and parking .
------------------------------------------------------------------------------------------------------------------------
......

Format of simulated dialogues

The simulated dialogues are saved in json format. For each dialogue, we save the following information:

  • dial_id: the id of the simulated dialogue, which consists of the ids of the used example dialogues. For example, comb_pmul3021_sng0548 is simulated with the examples of pmul3021 and sng0548.
  • goal: the user goal of this dialogue.
  • turns: a list of turns in this dialogue, where each turn is represented as a dictionary that contains the following fields:
    • dial_id - the unique ID for the dialogue session instance.
    • turn_num - this argument indicates the turn position in the dialogue session.
    • user - the user's utterance.
    • resp - the delexicalized reference system response.
    • bspn - the belief state.
    • aspn - the system action.
    • db - The database query result.
  • prompt: the prompt used to instruct GPT-3 to simulate the dialogue.

We provide the simulated dialogues in ./simulated_dialogues/ (w/o prompt for simplicity) and ./code/pptod/E2E_TOD/simulation_result23/small/ (w/ prompt) directory.

Turn-level simulation

You can use the following script to start simulating dialogue turns for DST augmentation.

cd ./pptod/E2E_TOD/
python dialogic_aug_dst.py\
  --train_data_ratio 0.01\
  --augment_time 2\
  --k_shot 2\
  --temperature 0.2

Demo

We also provide a demo to demonstrate how to simulate a dialogue turn by turn given a user goal. You can type into any user goal or use an automatically generated one to see how the corresponding dialogue is generated.

cd ./code/pptod/E2E_TOD/sh_folder/small/demo
chmod +x ./pptod_small_few_shot_0.01_demo.sh
./pptod_small_few_shot_0.01_demo.sh

An illustration of the demo example can be seen here.

Training on simulated dialogues

Convert the format of simulated dialogues for E2E training.

cd ./code/pptod/E2E_TOD/
python dialogic_post_process.py\
  --data_type E2E\
  --raw_data_path ./simulation_result23/small/few_shot_0.01/combine0.2_2_shot_augment_dialog_turn_info_train_ratio_0.01_simulation_result.json

Convert the format of simulated dialogue turns for DST training.

cd ./code/pptod/E2E_TOD/
python dialogic_post_process.py\
  --data_type DST\
  --raw_data_path ../data/multiwoz/data/multi-woz-2.3-dialogic-processed/2_shot_augment_x2_dst_turn_info_train_ratio_0.01.json

PPTOD

You can use the following scripts to train PPTOD:

# E2E
cd ./code/pptod/E2E_TOD/sh_folder/small/training/
chmod +x ./pptod_small_train_few_shot_0.01_augx1.sh
./pptod_small_train_few_shot_0.01_augx1.sh
# DST
cd ./code/pptod/DST/sh_folder/small/training/
chmod +x ./pptod_small_train_few_shot_0.01.sh
./pptod_small_train_few_shot_0.01.sh

SimpleTOD

Convert the format of simulated dialogues to fit SimpleTOD.

# E2E
cd ./code/pptod/E2E_TOD/
python dialogic_export_dialog_e2e.py\
  --train_data_ratio 0.01\
  --aug_train_data_file multi-woz-fine-processed-train-combine0.2_2_shot_augment_dialog_turn_info_train_ratio_0.01_simulation_result.json\
  --save_data_path_prefix ../../simpletod/resources_e2e_2.3_0.01_augx1/multi-woz
# DST
cd ./code/pptod/DST/
python dialogic_export_dialog_dst.py\
  --train_data_ratio 0.01\
  --aug_train_data_file multi-woz-fine-processed-train-2_shot_augment_x2_dst_turn_info_train_ratio_0.01.json\
  --save_data_path_prefix ../../simpletod/resources_DST_2.3_0.01_augx2/multi-woz

Then you can use the simulated dialogue to train SimpleTOD:

cd ./code/simpletod/
# create data
chmod +x create_dataset.sh
./create_dataset.sh
# E2E training
./train_end2end.sh 7 gpt2 gpt2 1
# DST training
./train_dst.sh 7 gpt2 gpt2 1

Use the following command for inference on test set:

CUDA_VISIBLE_DEVICES=$GPU python generate_dialogue_aug.py $CHECKPOINT $DECODING

Use the following command for evaluation:

python evaluate_multiwoz_aug.py $MODEL_OUTPUT $DATA_DIR

MinTL

You should use another environment for experiments on MinTL.

cd ./code/MinTL
pip install -r requirements.txt

Convert the format of simulated dialogues to fit MinTL.

# E2E
cd ./code/pptod/E2E_TOD/
python dialogic_export_dialog_e2e.py\
  --train_data_ratio 0.01\
  --aug_train_data_file multi-woz-fine-processed-train-combine0.2_2_shot_augment_dialog_turn_info_train_ratio_0.01_simulation_result.json\
  --save_data_path_prefix ../../MinTL/generated_data/e2e_2.3_0.01_augx1/
# DST
cd ./code/pptod/DST/
python dialogic_export_dialog_dst.py\
  --train_data_ratio 0.01\
  --aug_train_data_file multi-woz-fine-processed-train-2_shot_augment_x2_dst_turn_info_train_ratio_0.01.json\
  --save_data_path_prefix ../../MinTL/generated_data/dst_2.3_0.01_augx2/

Then you can use the simulated dialogue to train MinTL:

export PYTHONPATH='$PROJECT_PATH/code/MinTL/damd_multiwoz'
# E2E training
CUDA_VISIBLE_DEVICES=1 python train.py --mode train --context_window 2 --pretrained_checkpoint t5-small --cfg seed=557 batch_size=32 --use_db True --generated_data_file e2e_2.3_0.01_augx1
# DST training
CUDA_VISIBLE_DEVICES=1 python DST.py --mode train --context_window 3 --cfg seed=557 batch_size=32 --generated_data_file dst_2.3_0.01_augx2

Citations

If you found this repo useful, please consider citing our paper:

@article{li2022controllable,
  title={Controllable Dialogue Simulation with In-Context Learning},
  author={Li, Zekun and Chen, Wenhu and Li, Shiyang and Wang, Hong and Qian, Jing and Yan, Xifeng},
  journal={arXiv preprint arXiv:2210.04185},
  year={2022}
}

Acknowledgement

We thank the authors of SimpleTOD, MinTL, and PPTOD for sharing their code.

About

[EMNLP 2022] Code and data for "Controllable Dialogue Simulation with In-Context Learning"

Languages

Language:Python 97.3%Language:Shell 2.6%Language:Dockerfile 0.1%