XDAI

Paper | Slides

XDAI is maintained by the Knowledge Engineering Group of Tsinghua University, and supported by Zhipu.AI & Biendata, which is a tuning-free framework for exploiting pre-trained language models in knowledge grounded dialogue. Employing large-scale models including GLM and CPM, XDAI provides a basic open-domain dialogue generation service (XDAI-open) and several tools for developers in building domain-specific chatbots (XDAI-domain).

We summarize the features of XDAI as follows:

• Quick Start: XDAI provide an open-domain knowledge-grounded dialogue service with sufficient ready-to-use open-domain knowledge resources from XLore2. Developers can easily deploy a dialogue system with this basic service.
• Efficient Inference：Compared with other dialogue systems, XDAI utilizes a novel prompt pattern with knowledge injection, which optimizes the generated dialogues from PLMs without further training or tuning.
• Customized Deployment: For domain-specific developers XDAI provides easy-to-change plugins to enable automatically searching and updating the external knowledge only from a few domain-specific seeds.
• Incremental Modification: XDAI also provides a series of toolkits for incremental developing, encouraging developers to refine and customize their personalized components.

News (20221029) ‼️

• Our XDAI online service is conveniently available on https://tianqi.aminer.cn/open/document/cm_ref/xdai_chat !!

• Our paper "XDAI: A Tuning-free Framework for Exploiting Pre-trained Language Models in Knowledge Grounded Dialogue Generation" have been accepted by KDD2022 Applied Data Science track!!

• The XDAI service can be now experienced via our demo website. You can try to interact with different characters and DIY your own “小呆”（nickname of XDAI) and "take it home" from our Wechat Official Account: "小呆爱聊天":

Architecture

The overall architecture of XDAI is shown as bellow.

XDAI consists of two subsystems: online dialogue generation system & offline knowledge curation. Developers can employ their own toolkits and chatbot servers for local implementation.

Toolkit

XDAI Online API

You can apply for the just-to-use XDAI dialogue API via this document: https://tianqi.aminer.cn/open/document/cm_ref/xdai_chat

Directly deploy a chatbot and conduct scientific experiments with online api !!

PLM API (GLM)

You can deploy your own PLM server by taking the following steps. We offer the pack of GLM version in tools/PLM

1. Setup the environment (refer to https://github.com/THUDM/GLM).
2. Modify the [PLM]section in config/conf.ini:

   [PLM]
   ip_address = <server_ip>
   glm_port = <self-defined available port>
   glm_api = http://{ip_address}:{glm_port}/glm

   and set the "load_pretrained" in the tools/PLM/GLM/glm_config.json to be path-to-your-model.
3. Start the server by running:

   bash tools/deploy_plm.sh

4. If you use ready-made API, just set glm_api=<ready-made-PLM-api> and then add the corresponding function in module/use_plm

Knowledge Explore

While open-corpus is available, it is also encouraged to setup your own knowledge base.

The knowledge exploration toolkit at tools/knowledge is based on https://github.com/luogan1234/concept-expansion-snippet

Run the script to initiate model:

bash tools/init_knowledge_explore.sh

You can set the topic and seed concepts in data/seed_concept.json, and run the task:

# For init seed concepts
python tools/knowledge/explore.py -t init -f tools/knowledge/data/seed_concept.json
# For periodic knowledge exploration 
python tools/knowledge/explore.py -t update -f tools/knowledge/data/seed_concept.json -i 1 

The parameters are:

-t,--task:  init | update
-f,--config_file: config of the specific topic
-i,--interval: if the task is "update", this set the interval days between two consecutive updates.

FAQ toolkit

If you deploy your own knowledge exploration module as described above, you have to build a retrieval service.

We provide a faq toolkit based on fuzzybert which can be attached to the mongodb you use. Modify the faq_port in TOOL section in config/conf.ini and start the service:

bash tool/deploy_faq.sh

Sentence similarity

Modify the sentsim_port in TOOL section in config/conf.ini and start the service:

bash tool/deploy_sentsim.sh

You can also change the sbert-model referring to https://huggingface.co/models?library=sentence-transformers

QA-generation (based on T5)

If you want to use T5-based question generation instead of template-based, you have to build a generation service, which is based on https://github.com/patil-suraj/question_generation.

We provide a QA-gen toolkit, and you should modify the qagen_port in TOOL section in config/conf.ini and start the service:

bash tool/deploy_t5QA.sh.sh

Notice that transformer==3.0.0 is required by the source code. This question generation code only supports English question generation, so we use youdao translation for English-Chinese translation, which is implemented in utils/translate.py. This translation interface has a sending limit. If you need to use it, please replace it with your own translation interface.

Get Started

1. Requirements

pip install -r requirements.txt

2. Set up PLM API

Get a ready-made PLM generation api or deploy one with the toolkit mentioned above.

3. Design your Bot Agent

An agent class corresponds to an independent processing flow of prompt manufacturing, which should be derived from the base class in agents/agent_base.py:

from agents import AgentBase

class YourOwnAgent(AgentBase):
    ...

You can add agent classes in agents/. There are three derived subclasses as examples.

agents/
├── agent_base.py
├── __init__.py
├── xdai_glm.py
├── xdai_kg.py
└── xdai_kg_specific.py

• xdai_glm.py: The baseline implementation without knowledge injected using GLM as the PLM.
• xdai_kg.py: GLM + open knowledge injected using XLore2
• xdai_kg_specific.py:GLM + specific domain knowledge with self-maintained FAQ db.

After being added, a new class should then be imported in the agents/__init__.py and appended to IN_USE_AGENTS

4. Interact

The following 3 methods of deployment are offered to interact with the chatbot:

1. Terminal: The simplest way to expeirence the dialogue system, which does not require the api server as the prerequisite.

bash scripts/run_terminal_chat.sh

2. Streamlit: It is started with some backend apis deployed. Run the bash:

bash scripts/run_streamlit.sh

For more information & instruction : https://docs.streamlit.io/

3. Wechaty:

Link the chatbot to your wechat account, with backend server required as well. Before start, you need to get a token from http://wechaty.js.org/ and set the token in conf.ini. Then you can run:

bash scripts/run_wechaty.sh

Analysis & Example

XDAI can achieve competitive results in both open-domain and domain-specific dialogue scenarios. The table shows 95% confidence intervals for human evaluation results of open-domain dialogue generation.

There is a case of how the explored knowledge help lifting the informativeness of generated dialogues.

Reference

    @inproceedings{10.1145/3534678.3539135,
        author = {Yu, Jifan and Zhang, Xiaohan and Xu, Yifan and Lei, Xuanyu and Guan, Xinyu and Zhang, Jing and Hou, Lei and Li, Juanzi and Tang, Jie},
        title = {XDAI: A Tuning-Free Framework for Exploiting Pre-Trained Language Models in Knowledge Grounded Dialogue Generation},
        year = {2022},
        isbn = {9781450393850},
        publisher = {Association for Computing Machinery},
        address = {New York, NY, USA},
        url = {https://doi.org/10.1145/3534678.3539135},
        doi = {10.1145/3534678.3539135},
        booktitle = {Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
        pages = {4422–4432},
        numpages = {11},
        keywords = {dialogue systems, pre-trained model exploitation},
        location = {Washington DC, USA},
        series = {KDD '22}
    }