Plug-and-Play Conversational Models

This is the implementation of the paper:

Plug-and-Play Conversational Models. Andrea Madotto, Etzuko Ishii, Zhaojiang Lin, Sumanth Dathathri, Pascale Fung [PDF] EMNLP2020 (findings)

If you use any source codes or datasets included in this toolkit in your work, please cite the following paper. The bibtex is listed below:

@inproceedings{madotto2020plug,
  title={Plug-and-Play Conversational Models},
  author={Madotto, Andrea and Ishii, Etsuko and Lin, Zhaojiang and Dathathri, Sumanth and Fung, Pascale},
  booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings},
  pages={2422--2433},
  year={2020}
}

Abstract

There has been considerable progress made towards conversational models that generate coherent and fluent responses; however, this often involves training large language models on large dialogue datasets, such as Reddit. These large conversational models provide little control over the generated responses, and this control is further limited in the absence of annotated conversational datasets for attribute specific generation that can be used for fine-tuning the model. In this paper, we first propose and evaluate plug-and-play methods for controllable response generation, which does not require dialogue specific datasets and does not rely on fine-tuning a large model. While effective, the decoding procedure induces considerable computational overhead, rendering the conversational model unsuitable for interactive usage. To overcome this, we introduce an approach that does not require further computation at decoding time, while also does not require any fine-tuning of a large language model. We demonstrate, through extensive automatic and human evaluation, a high degree of control over the generated conversational responses with regard to multiple desired attributes, while being fluent.

Plug-and-Play Conversational Models (PPCM)

Basic Usage

Dependencies

Create a python3.6 virtual environment and run pip install -r requirements.txt.

Discriminator Training

python dialogGPT_discr.py --save_model --dataset sentiment --cached --epochs 100 
python dialogGPT_discr.py --save_model --dataset daily_dialogue_act --cached --epochs 100 
python dialogGPT_discr.py --save_model --dataset TC_AG_NEWS --cached --epochs 50 
python dialogGPT_discr.py --save_model --dataset TC_SogouNews --cached --epochs 50 
python dialogGPT_discr.py --save_model --dataset TC_DBpedia --cached --epochs 10 
python dialogGPT_discr.py --save_model --dataset TC_YahooAnswers --cached --epochs 10 

Run PPLM

By omitting the --evaluate flag, you can run PPLM in an interactive mode.

python main.py -D AG_NEWS --label_class 0 --length 30 --num_samples 10 --evaluate --verbose --all_starter --wd
python main.py -D AG_NEWS --label_class 1 --length 30 --num_samples 10 --evaluate --verbose --all_starter --wd
python main.py -D AG_NEWS --label_class 2 --length 30 --num_samples 10 --evaluate --verbose --all_starter --wd
python main.py -D AG_NEWS --label_class 3 --length 30 --num_samples 10 --evaluate --verbose --all_starter --wd
python main.py -D sentiment --label_class 3 --length 30 --num_samples 10 --evaluate --verbose --all_starter
python main.py -D sentiment --label_class 2 --length 30 --num_samples 10 --evaluate --verbose --all_starter
python main.py -D daily_dialogue_act --label_class 1 --length 30 --num_samples 10 --evaluate --verbose --all_starter

Run Adapter

python train_supervised_adapter.py --dataset SENT --label very_negative --iter 75 --lr 6.25e-4
python train_supervised_adapter.py --dataset SENT --label very_positive --iter 25
python train_supervised_adapter.py --dataset QUEST --label question --iter 25
python train_supervised_adapter.py --dataset TOPI --label Business --iter 25
python train_supervised_adapter.py --dataset TOPI --label SciTech --iter 25
python train_supervised_adapter.py --dataset TOPI --label Sports --iter 25

Reproducibility

You can simply run ./download_data.sh to download and extract all required files, or you can perform the required actions manually, by following the steps outlined bellow:

Manual setup

Dataset

Download the datasets

❱❱❱ unzip data.zip

DialoGPT

Download dialoGPT

❱❱❱ unzip dialoGPT.zip
❱❱❱ mv dialiGPT models

Discriminators

Download the discriminators

❱❱❱ unzip discriminators.zip
❱❱❱ mv discriminators models

Scorers

Download the scorers

❱❱❱ unzip scorers.zip
❱❱❱ mv scorers models

Reproducibility

Download the generated responses

❱❱❱ unzip evaluate.zip
❱❱❱ mv evaluate results
❱❱❱ python evaluate.py

In each folder you can find response generated by PPLM using multple styles. This are use to train each of the adapter using train_supervised_adapter.py.

Download the Human Evaluation Score

❱❱❱ unzip human_evaluation.zip

here you can find the jupiter notebook to replicate the human evaluation results and the human judgment scores.

Run Check the experiment_runner folder to see how to run the generation.

Toxicity

For researcher working on abusive language, we have also responses generated using a toxic classifer. We can release it upon request exclusively for research purposes.

Acknowledgement

We would like to thanks the MLC for the feedback on the earlystage of the work, and expecially Jason Yosinski. This repository is implemented base on Huggingface