Awesome

<img src=figure/icon.png width=8% />Pchatbot: A Large-Scale Dataset for Personalized Chatbot

Introduction

We introduce Pchatbot, a large scale conversation dataset dedicated for the development of personalized dialogue models. In this dataset, we assign anonymized user IDs and timestamps to conversations. Users’ dialogue histories can be retrieved and used to build rich user profiles. With the availability of the dialogue histories, we can move from personality based models to personalized models.

Pchatbot has two subsets, named PchatbotW and PchatbotL, built from open-domain Weibo and judicial forums respectively. Since the data volume of each sub-data set is too large, we divided each sub-data set into 10 equal parts according to the number of users, and named them PchatbotW-i and PchatbotL-i.

The dataset paper is accepted to SIGIR 2021 (Resource Track). See paper for more details.

Citation

If you use the dataset in your work, please cite:

@inproceedings{qian2021pchatbot,
     author = {Hongjin Qian and Xiaohe Li and Hanxun Zhong and Yu Guo and Yueyuan Ma and Yutao Zhu and Zhanliang Liu and Zhicheng Dou and Ji-Rong Wen}, 
     title = {Pchatbot: A Large-Scale Dataset for Personalized Chatbot}, 
     booktitle = {Proceedings of the {SIGIR} 2021}, 
     publisher = {{ACM}}, 
     year = {2021}, 
     url = {https://doi.org/10.1145/3404835.3463239}, 
     doi = {10.1145/3404835.3463239}}

The following paper uses the Pchatbot dataset:

1. One Chatbot Per Person: Creating Personalized Chatbots based on Implicit User Profiles (SIGIR 2021 Long Paper)

@inproceedings{DBLP:conf/sigir/madousigir21,
     author = {Zhengyi Ma and Zhicheng Dou and Yutao Zhu Hanxun Zhong and Ji-Rong Wen}, 
     title = {One Chatbot Per Person: Creating Personalized Chatbots based onImplicit User Profiles}, 
     booktitle = {Proceedings of the {SIGIR} 2021}, 
     publisher = {{ACM}}, 
     year = {2021}, 
     url = {https://doi.org/10.1145/3404835.3462828}, 
     doi = {10.1145/3404835.3462828}}

2. Learning Implicit User Profile for Personalized Retrieval-based Chatbot (CIKM 2021 Long Paper)

@inproceedings{qian2021impchat,
     author = {Hongjin Qian and Zhicheng Dou and Yutao Zhu Yueyuan Ma and Ji-Rong Wen}, 
     title = {Learning Implicit User Profile for Personalized Retrieval-based Chatbot}, 
     booktitle = {Proceedings of the {CIKM} 2021}, 
     publisher = {{ACM}}, 
     year = {2021},
     url = {https://doi.org/10.1145/3459637.3482269},
     doi = {10.1145/3459637.3482269}

3. Less is More: Learning to Refine Dialogue History for Personalized Dialogue Generation (NAACL 2022 Long Paper)

@inproceedings{zhong-etal-2022-less,
    title = "Less is More: Learning to Refine Dialogue History for Personalized Dialogue Generation",
    author = "Zhong, Hanxun  and
      Dou, Zhicheng  and
      Zhu, Yutao  and
      Qian, Hongjin  and
      Wen, Ji-Rong",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.426",
    doi = "10.18653/v1/2022.naacl-main.426",
    pages = "5808--5820"

Dataset Statistics

The detailed statistics of Pchatbot shows as follow:

We construct two standard dataset from Pchatbot for both generation-based and retrieval-based tasks, named PchatbotW-R and PchatbotW-G. The two datasets can be directly used in coressponding dialogue tasks. We will release the standard dataset then. Their statistics are shown in the following table:

To obtain statistics, run:

python src/statistics.py

We will then release standard datasets for PchatbotL.

Data Content and Format

Obtain the data

For now we provide download link via Baidu Cloud which may be a bit slow outside mainland, China. We will update Google drive link asap.

Pchatbot-L:

md5: 48bd7ab93f625ebdf34c7254ff27ac2a

Pchatbot-W:

md5: cd443951973f47f5614df298e6e416da

If you cannot access Baidu Cloud Disk, contact us and we will try to provide other options.

Please fill in the application form and send it to the contact mail, we will then send the download links and the password for Baidu Cloud Disk to you. Note that the application form should be signed by the person in charge of your research group. We will update the download password regularly.

Application Form

Pchatbot Files

The upload format of the dataset is .tar.bz2, you can decompress it as follows：

tar -jxvf xx.tar.bz2

The format of each piece of data in the data set is：

Post \t Post_user_id \t Post_timestamp \t Response \t Response_user_id \t Response_timestamp \n

post and response are sentences with word segmentation, separated by spaces.And we give several examples of the data in data/sample.txt

We also give some examples of user personalized information, as shown in the figure below, due to space constraints, we only selected 5 historical records for the user in each example. PchatbotW.release_ver

Data Preprocessing

Instructions for data cleaning, preprocessing, aggregation and dataset constructs are in ./src/ folder.

Baseline models

We provide results of baseline models on the PchatbotW-R and PchatbotW-G dataset. For evaluation details, please refer to our paper. We will continue to update the results of other baseline models:

PchatbotW-R

PchatbotW-G

License

This repository is liciensed under Apache-2.0 License.

The Pchatbot dataset is liciensed under CC BY-NC 2.0.

FAQ