File size: 2,725 Bytes
3f86df3 83a1944 a529f65 be1bac8 83a1944 3f86df3 a529f65 f8c6b17 3f86df3 f8c6b17 e0cf98b 3f86df3 83a1944 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 |
# DialogRPT-updown
### Dialog Ranking Pretrained Transformers
> How likely a dialog response is upvoted 👍 and/or gets replied 💬?
This is what [**DialogRPT**](https://github.com/golsun/DialogRPT) is learned to predict.
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.
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.
Quick Links:
* [EMNLP'20 Paper](https://arxiv.org/abs/2009.06978/)
* [Dataset, training, and evaluation](https://github.com/golsun/DialogRPT)
* [Colab Notebook Demo](https://colab.research.google.com/drive/1cAtfkbhqsRsT59y3imjR1APw3MHDMkuV?usp=sharing)
We considered the following tasks and provided corresponding pretrained models. This page is for the `updown` task, and other model cards can be found in table below.
|Task | Description | Pretrained model |
| :------------- | :----------- | :-----------: |
| **Human feedback** | **given a context and its two human responses, predict...**|
| `updown` | ... which gets more upvotes? | (this model) |
| `width`| ... which gets more direct replies? | [model card](https://huggingface.co/microsoft/DialogRPT-width) |
| `depth`| ... which gets longer follow-up thread? | [model card](https://huggingface.co/microsoft/DialogRPT-width) |
| **Human-like** (human vs fake) | **given a context and one human response, distinguish it with...** |
| `human_vs_rand`| ... a random human response | [model card](https://huggingface.co/microsoft/DialogRPT-human-vs-rand) |
| `human_vs_machine`| ... a machine generated response | [model card](https://huggingface.co/microsoft/DialogRPT-human-vs-machine) |
### Examples:
The `updown` score predicts how likely the response is getting upvoted.
Examples below can be reproduced with this [Colab Notebook](https://colab.research.google.com/drive/1cAtfkbhqsRsT59y3imjR1APw3MHDMkuV?usp=sharing)
| Context | Response | `updown` score |
| :------ | :------- | :------------: |
| I love NLP! | Here’s a free textbook (URL) in case anyone needs it. | 0.613 |
| I love NLP! | Me too! | 0.111 |
### Contact:
Please create an issue on [our repo](https://github.com/golsun/DialogRPT)
### Citation:
```
@inproceedings{gao2020dialogrpt,
title={Dialogue Response RankingTraining with Large-Scale Human Feedback Data},
author={Xiang Gao and Yizhe Zhang and Michel Galley and Chris Brockett and Bill Dolan},
year={2020},
booktitle={EMNLP}
}
```
|