Update README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# DialogRPT-width
|
| 2 |
+
|
| 3 |
+
### Dialog Ranking Pretrained Transformers
|
| 4 |
+
|
| 5 |
+
> How likely a dialog response is upvoted 👍 and/or gets replied 💬?
|
| 6 |
+
|
| 7 |
+
This is what [**DialogRPT**](https://github.com/golsun/DialogRPT) is learned to predict.
|
| 8 |
+
It is a set of dialog response ranking models proposed by [Microsoft Research NLP Group](https://www.microsoft.com/en-us/research/group/natural-language-processing/) trained on 100 + millions of human feedback data.
|
| 9 |
+
It can be used to improve existing dialog generation model (e.g., [DialoGPT](https://huggingface.co/microsoft/DialoGPT-medium)) by re-ranking the generated response candidates.
|
| 10 |
+
|
| 11 |
+
Quick Links:
|
| 12 |
+
* [EMNLP'20 Paper](https://arxiv.org/abs/2009.06978/)
|
| 13 |
+
* [Dataset, training, and evaluation](https://github.com/golsun/DialogRPT)
|
| 14 |
+
* [Colab Notebook Demo](https://colab.research.google.com/drive/1cAtfkbhqsRsT59y3imjR1APw3MHDMkuV?usp=sharing)
|
| 15 |
+
|
| 16 |
+
We considered the following tasks and provided corresponding pretrained models.
|
| 17 |
+
|
| 18 |
+
|Task | Description | Pretrained model |
|
| 19 |
+
| :------------- | :----------- | :-----------: |
|
| 20 |
+
| **Human feedback** | **given a context and its two human responses, predict...**|
|
| 21 |
+
| `updown` | ... which gets more upvotes? | [model card](https://huggingface.co/microsoft/DialogRPT-updown) |
|
| 22 |
+
| `width`| ... which gets more direct replies? | [model card](https://huggingface.co/microsoft/DialogRPT-depth) |
|
| 23 |
+
| `depth`| ... which gets longer follow-up thread? | this model |
|
| 24 |
+
| **Human-like** (human vs fake) | **given a context and one human response, distinguish it with...** |
|
| 25 |
+
| `human_vs_rand`| ... a random human response | [model card](https://huggingface.co/microsoft/DialogRPT-human-vs-rand) |
|
| 26 |
+
| `human_vs_machine`| ... a machine generated response | [model card](https://huggingface.co/microsoft/DialogRPT-human-vs-machine) |
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
### Examples:
|
| 31 |
+
The `depth` score predicts how likely the response is getting a long follow-up discussion thread.
|
| 32 |
+
Examples below can be reproduced with this [Colab Notebook](https://colab.research.google.com/drive/1cAtfkbhqsRsT59y3imjR1APw3MHDMkuV?usp=sharing)
|
| 33 |
+
|
| 34 |
+
| Context | Response | `depth` score |
|
| 35 |
+
| :------ | :------- | :------------: |
|
| 36 |
+
| I love NLP! | Can anyone recommend a nice review paper? | 0.701 |
|
| 37 |
+
| I love NLP! | Me too! | 0.029 |
|
| 38 |
+
|
| 39 |
+
### Contact:
|
| 40 |
+
Please create an issue on [our repo](https://github.com/golsun/DialogRPT)
|
| 41 |
+
|
| 42 |
+
### Citation:
|
| 43 |
+
```
|
| 44 |
+
@inproceedings{gao2020dialogrpt,
|
| 45 |
+
title={Dialogue Response RankingTraining with Large-Scale Human Feedback Data},
|
| 46 |
+
author={Xiang Gao and Yizhe Zhang and Michel Galley and Chris Brockett and Bill Dolan},
|
| 47 |
+
year={2020},
|
| 48 |
+
booktitle={EMNLP}
|
| 49 |
+
}
|
| 50 |
+
```
|