Skip to content

Latest commit

 

History

History
236 lines (200 loc) · 12 KB

README_EN.md

File metadata and controls

236 lines (200 loc) · 12 KB
Raw

GLM-Open-Dialogue

A enhanced Open Dialogue Context Generator supported by General Language Model Pretraining with Autoregressive Blank Infilling

中文介绍

Brief introduction

Introduction and related works

Given the definition of Open-Domain Dialogue Generation Open-domain Dialogue Generation: What We Can Do, Cannot Do, And Should Do Next

Given zero or more previous dialogue turns between
itself and one or more other participants,
a system  output a fluent, engaging,
and meaningful natural language response.

A well known model deal with this task is DialoGPT, You can try DialoGPT on Roxza DialoGPT space

And you can also take look at my repository svjack/Daliy-Dialogue to try Bloom or GPT model on this kind of task.

Dataset generate by models in svjack/Daliy-Dialogue demonstration :

Name HuggingFace Dataset link HuggingFace Space link Language
English Daliy Dialogue model generate samples 🦅🌸 https://huggingface.co/datasets/svjack/bloom-dialogue-generate-ds-en https://huggingface.co/spaces/svjack/bloom-dialogue-english-sample-search English
Chinese Dialogue model generate samples 🐰🌸 https://huggingface.co/datasets/svjack/bloom-dialogue-generate-ds-zh https://huggingface.co/spaces/svjack/bloom-gpt-dialogue-chinese-sample-search Chinese

English Examples of this project

Following is a English example compare the output of DialoGPT and GLM(this project do)

Model First Question Response chain
DialoGPT Does money buy happiness? Depends how much money you spend on it.
What is the best way to buy happiness ?
You just have to be a millionaire by your early 20s, then you can be happy .
This is so difficult !
You have no idea how hard it is to be a millionaire and happy . There is a reason the rich have a lot of money
GLM Does money buy happiness? Can money buy happiness?
Yes, what about you?That is happiness.
Oh yeah!
You will be happy.
Oh, I got it.How are you happy?
I think no.
Everyone has different definitions of happiness, and money cannot buy happiness.
DialoGPT what is the meaning of life ? The meaning is to be with your family and friends .
I’m going to guess : It means that your parents and friends have loved you . That’s the meaning , I believe .
I think the meaning is to make sure you’re alive when you die . So you’ll be with your family , friends , and yourself .
It’s not , but thanks for trying to help .
GLM what is the meaning of life ? You just like life, answer?
of course is.
Who said life itself?
Does life have no meaning?
Yes is.
What do you like to do with me that is meaningful?
Enjoy every moment of your life and feel meaningful.

HuggingFace Space demonstration

Installation and Instructions

pip install -r req0.txt
pip install -r req1.txt

Instructions

I recommand run the demo on a computer with gpu: If you have gpu: change the first line of predict_t5_choose_simple.py from

#### Run
#### pip install torch==1.8.0
device = "cpu"

to

#### Run
#### replace by your cuda version.
#### pip install torch==1.8.0+cu111
device = "cuda:0"

Then

from main import *

#### English Usage
##### max_times set to 10 will take more times than 2, and more outputs 
demo_func(*["Which keyboard do you prefer?", "en", 2])
demo_func(*["Which keyboard do you prefer?", "en", 10])
#### Japanese Usage
demo_func(*["東京と京都のどちらが住みやすいですか。", "ja", 2])
demo_func(*["東京と京都のどちらが住みやすいですか。", "ja", 10])

#### Chinese Usage
demo_func(*["程序员要掌握哪些技能", "zh", 5])
### OR
generate_func(*["程序员要掌握哪些技能", 5])

generate_func(*["你对《三国演义》感兴趣吗?", 10])
generate_func(*["谁能让美国再一次伟大?", 10])

Construction

Migrate GLM to Open-Domain Dialogue Generation domain

Many this kinds models focus train a generate model on some specific datasets, This may constraints from datasets and pretrained models, When faced with a open domain with a broad breadth of knowledge, the generator may has more risk of cause Blandness Problem by bland answers or give some response lack of vividness.
This project deal with this kinds of problems by a unsupervised combination method. And decompose this problem to four steps:

  • 1 Self-train a Open-Domain Dialogue generator on a small dataset from scratch and use it as a "Construct Generator"
  • 2 Use FlagAI's GLM as "Knowledge Generator"
  • 3 Use a self-trained "Context Discriminer" to combine above two generators.
  • 4 Use a self-trained "Context Reconstructor" to reconstruct the context.

There is no high requirement for the accuracy of the "Construct Generator".

The above four models are released in HuggingFace, list in the below sheet.

Role in Project HuggingFace link Contributor
Construct Generator https://huggingface.co/svjack/T5-daliy-dialogue https://huggingface.co/svjack
Knowledge Generator https://huggingface.co/BAAI/glm-large-chinese https://huggingface.co/BAAI
Context Discriminer https://huggingface.co/svjack/T5-dialogue-choose https://huggingface.co/svjack
Context Reconstructor https://huggingface.co/svjack/T5-dialogue-collect https://huggingface.co/svjack

They all build above Chinese, and use facebook m2m_100_1.2B translator make them support Multilanguage domain.

Prompt --- basic concept throughout the project

There are four entry points Prompt make for this project.

Prompt category Effect parts Idea source Prompt format in English Prompt format in Chinese
Task Oriented Prompt Format for QA Blank Infilling GLM build in GLM Question: Is reading good for your grow up? Answer: [MASK] 问题: 阅读对于成长有好处吗？ 回答: [MASK]
Task Oriented Prompt Format for choose best Context Discriminer Bigscience-mt0 According to the following context, choose the optimal follow-up sentence Context: {}\n{}Answer: 根据如下上下文，选择最优的后续句子 上下文：{}\n{}答案：
Task Oriented Prompt Format for reconstruct context Context Reconstructor Bigscience-mt0 According to the following context, reconstruct the Context: {}\n{}Answer: 根据下面的上下文进行分段： 上下文：{} 答案：
Quality Oriented Negative Prompt negative_prompts in predict_t5_choose_simple.py Stable Diffusion [Yes, I don't know, Who knows] [是的,怎么知道]

Examples

Language First Question Response chain
English Which keyboard do you like? My true love!
My favorite mechanical keyboard!
The mechanics.keyboard
HP Mechanical Keyboard
Of course a mechanical keyboard.
I like the mechanical keyboard, though not the hippie mouse.
I like it, but listen.Do you want?
HP's mechanical keyboard is really useful, and the cost is very high, but
Japanese 東京と京都はどちらがいいですか? はい。,..わからないけど、あなたはどうですか?
京都!
うん、確信がない。,..あなたはどうですか?
京都に行くと東京は観光都市です。
誰もが京都よりはマシだと思っている。また、日本の都市
個人的には京都の方がお得だと思います。
知っている。あなたはエッフェル塔の頂上に行ったことがありますか?
Chinese 你对《三国演义》感兴趣吗? 是的,我喜欢各种人物。你觉得怎么样?
我觉得可以,因为太无聊了。我喜欢《三国演义》,但是没时间去看。
抖机灵,我喜欢历史,《三国志》。这些经典读起来不枯燥,有很强的吸引力,很适合初学者。
哦,我也喜欢历史。你有兴趣吗?
看历史,我感兴趣。
毕竟,《三国志》是经典,但是要看清人物才能更好地理解人物性格。

Similar but more academic work

There is a work named "Controllable Factuality in Document-Grounded Dialog Systems Using a Noisy Channel Model" by UKPLab
Paper link: https://arxiv.org/pdf/2210.17418.pdf
Github link: https://github.com/ndaheim/noisy_channel_model
It's also about some sense of generate and "fix" on Dialog Systems, The mainly difference between that work and mine is,

It "decode and reconstruct" tokens rather than my "choose and reconstruct" sentence.
It use SentenceTransformer to recall samples and i use GLM to "recall" knowledge.

If you are interesting in that more cademic work, I suggest you have a try.

Sample Dataset Generated

A sample dataset generated in Chinese has been uploaded to Huggingface Hub.
GLM-Open-Dialogue-Chinese-samples
The questions in this dataset may have some difficulties for extractive question answering systems which provide context. If you try some other questions(harder or simple), the outputs may be different.

Contact

svjack - https://huggingface.co/svjack - svjackbt@gmail.com - ehangzhou@outlook.com

Project Link:https://github.com/svjack/GLM-Open-Dialogue

Acknowledgements