This repo contains the dataset from NAACL 2022 paper "KETOD: Knowledge-Enriched Task-Oriented Dialogue" https://arxiv.org/abs/2205.05589
Notes: I made this modification to works with latest version of pytorch and huggingface transformers.
Also, I provided the simple_demo.py in simpletodplus folder to run the model with the trained model.
.
Make sure to install Pytorch https://pytorch.org/get-started/locally and another dependencies packages
pip install -r requirements.txt
or you can use environment.yml preferable for anaconda environment
If you never heard about Google SGD dataset, please read it here: https://blog.research.google/2019/10/introducing-schema-guided-dialogue.html
KETOD is built upon the google SGD dataset and combined with author knowledge-enriched utterances annotations and the script to generate the final dataset.
To generate the full KETOD dataset in root folder, execute this:
python gen_ketod_data.py
Notes: this will combine between SGD dataset and enriched knowledge dataset from author.
Three new files generated at ketod_release folder: train_final.json, dev_final.json, and test_final.json.
Have a look on my detailed comments in the file gen_ketod_data.py or see the dataset format below.
Each entry of the data is one dialogue. It has the following fields:
"dialogue_id": unique id of the dialogue.
"turns": the list of dialogue turns. Besides the original fields in the SGD dataset, if it is an enriched turn, then we have the following additional fields:
{
"enrich": True. For turns without chitchat enrichment, this field is False.
"entity_query": The entity query we use to do knowledge retrieval.
"enriched_utter": The utterance enriched with chitchat. Another field 'utterance' is the original response in the SGD dataset.
"kg_snippets": the index of the ground truth knowledge snippets
"kg_snippets_text": the content of the ground truth knowledge snippets
}
"dialog_query": all the entity queries we use to do knowledge retrieval in this dialog
"entity_passages": all the wikipedia passages retrieved in this dialog
"entity_passage_sents": all the wikipedia passages retrieved in this dialog, breaked into snippets associated with index numbers
To run the knowledge selection model, go to "kg_selection" folder and execute process_data.py
Have a look my detailed comments inside the file
cd code
python process_data.py --data train_final.json
python process_data.py --data dev_final.json
python process_data.py --data test_final.json
And this will produce file processed_kg_select_train_final.json, processed_kg_select_dev_final.json and processed_kg_select_test_final.json in ketod_release folder
Edit file run_model.sh and adjust your GPU. Next, edit config_py and adjust the batch_size depending on your GPU.
Train the model
bash run_model.sh
You will see on the outputs folder the model is saved. For example: kg_select_bert_base__2023....
Modify saved_model_path in config.py to refer into the model checkpoint. For example:
saved_model_path = output_path + "kg_select_bert_base__20231026164959/saved_model/loads/10/model.pt"
Generate train, valid and test prediction.json using Test.py
Modify the Test.py by comment and uncomment the following lines to use the train, val and test data. For example:
# modify this line to use the train, val and test data
test_data, test_examples = read_examples(input_path=conf.train_file, is_inference=True)
# test_data, test_examples = read_examples(input_path=conf.valid_file, is_inference=True)
# test_data, test_examples = read_examples(input_path=conf.test_file, is_inference=True)
Un-comment Test.py in run_model.sh, and comment Main.py
Run bash run_model.sh and keep modify it to produce train``, valid`` and test.
bash run_model.sh
This will generate inference output, for example: inference_only_20231023190626_kg_select_bert_base_/results/test/predictions.json
To run the SimpleToDPlus model, go to simpletodplus, modify and run gen_kg_train.py to generate data files with the kg selection results.
# test
json_in = root + "inference_only_20231023190626_kg_select_bert_base_/results/test/predictions.json"
json_out = root + "model1/" + "test_retrieved.json"
merge_train(json_in, json_out, topn_snippets=3, is_test=True)
Then run gen_data.py to generate train/dev/test files for the model input formats.
Using the run_simpletod.sh script, run train_simpletod.py for training, and test_simpletod_simple.py for testing. You need to modify and follow the steps at the end of the test_simpletod_simple.py file to generate the results for each step.
To run the demo model, go to simpletodplus, run simple_demo.py <path_to_checkpoint>.
For example:
python simple_demo.py ../../outputs/model1_rand_20231027120607/saved_model/loads/21/model.pt
If you find this project useful, please cite it using the following format
@inproceedings{DBLP:conf/naacl/ChenLMSCW22,
author = {Zhiyu Chen and
Bing Liu and
Seungwhan Moon and
Chinnadhurai Sankar and
Paul A. Crook and
William Yang Wang},
editor = {Marine Carpuat and
Marie{-}Catherine de Marneffe and
Iv{\'{a}}n Vladimir Meza Ru{\'{\i}}z},
title = {{KETOD:} Knowledge-Enriched Task-Oriented Dialogue},
booktitle = {Findings of the Association for Computational Linguistics: {NAACL}
2022, Seattle, WA, United States, July 10-15, 2022},
pages = {2581--2593},
publisher = {Association for Computational Linguistics},
year = {2022},
url = {https://doi.org/10.18653/v1/2022.findings-naacl.197},
doi = {10.18653/v1/2022.findings-naacl.197},
timestamp = {Mon, 01 Aug 2022 16:27:57 +0200},
biburl = {https://dblp.org/rec/conf/naacl/ChenLMSCW22.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
KETOD is released under MIT license, see LICENSE for details.