automatic-data-generation

Pytorch implementation of experiments described in "Conditioned Text Generation with Transfer for Closed-Domain Dialogue Systems" by S. d'Ascoli, A. Coucke, F. Caltagirone, A. Caulier, M. Lelarge accepted for publication at the 8th International Conference on Statistical Language and Speech Processing (2020). This is a work in progress, feel free to reach out for any question.

Install

Requirements: Python3.6, pip

virtualenv venv
. venv/bin/activate
pip install -e .

You might need to download some NLTK resources:

  >>> import nltk
  >>> nltk.download('punkt')

Dataset

The abstract class automatic-data-generation.data.base_dataset.py provides the interface for representing a training dataset. To implement a new dataset format, write a class inheriting from Dataset and implement its abstract methods.

You then need to allow for a new dataset_type in the training script automatic_data_generation/train_and_eval_cvae.py which should be the name of the sub-directory in your data folder.

Finally, you should update the data factory create_dataset in automatic_data_generation/utils/utils.py.

None sentences

The reservoir dataset of unannotated queries used for transfer experiments in the paper is not publicly available. To explore the query transfer method, you need to add you own None sentences in csv format in a sub-directory in your data folder.

You need to first download the InferSent model by running the following executable:

./automatic_data_generation/data/get_infersent.sh

To embed your None sentences, run the following command:

python automatic_data_generation/data/utils/embed_intents.py --dataset_path 
./your/none/data/path

You then need to allow for a new none_type in the training script automatic_data_generation/train_and_eval_cvae.py which should be the name of the sub-directory in your data folder.

Add the index of the utterances in the csv file to the NONE_COLUMN_MAPPING dictionary in automatic_data_generation/data/utils/utils.py. If the utterance is the first (resp. n) field of the csv, add 0 (resp. n+1).

You should be good to go.

Training

Use the script automatic-data-generation.train_and_eval_cvae.py to train a model, generate sentences, and evaluate their quality.

For a simple run without query transfer, you may run:

python automatic_data_generation/train_and_eval_cvae.py -ep 10 --n-generated 100 --dataset-size 125

Possible options are:

• --dataset-size: number of sentences in the training dataset
• --none-size: number of None sentences to be added to the training dataset
• --none-type: type of None sentences
• --restrict-to-intent: list of intents to filter on for training
• --n-epochs: number of epochs for training
• --n-generated: number of generated sentences
• --infersent-selection: possible query transfer schemes, unsupervised is the normal scheme, supervised is the pseudolabelling baseline, and NO_INFERSENT_SELECTION deactivates the feature
• --cosine-thresholds: the selection threshold for query transfer (defaults to 0.9)
• alpha: the parameter regulating transfer

If you have added your own None type, a typical run may be:

python automatic_data_generation/train_and_eval_cvae.py -ep 50 --n-generated
 2000 --dataset-size 125 --none-size 125 --none-type mynonetype 
 --infersent-selection unsupervised --cosine-threshold 0.9 --alpha:0.1

Output folder

An folder will be created with the following elements:

• load: a folder with a model.pth file and its associated config.json and a vocab.pth file containing the vocabulary
• tensorboard: a folder with the checkpoints for tensorboard
• run.pkl: a dictionnary with every runtime parameters
• train_*.csv: the training dataset
• train_*_augmented.csv: the training dataset augmented with generated sentences
• validate_*.csv: the validation dataset