Continual Learning for Natural Language Generation in Task-oriented Dialog Systems

Introduction

This repo is the implementation of paper Continual Learning for Natural Language Generation in Task-oriented Dialog Systems on Findings of ACL: EMNLP 2020. In this paper, we propose a exemplar-replay and regularization based technique to mitigate catastrophic forgetting and enhance continual learning for language generation in task oriented dialog system.

Authors:

Fei Mi, Liangwei Chen, Mengjie Zhao, Minlie Huang, Faltings Boi.

Prerequisites

The following modules are necessary to run the experiments

• nltk 3.4.5 conda install -c anaconda nltk
• PyTorch 1.3.1 conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
• pandas 0.20.3 conda install -c anaconda pandas
• numpy 1.11.3 conda instal -c anaconca numpy

Code structure

.
├── model/                      # The model implementation containing the sclstm layers  
│  ├── lm_deep.py               # SCLSTM basedimplementation of utterance generation
│  ├── cvae.py                  # SCVAE based implementation of utterance generation
│  ├── masked_cross_entropy.py  # Masked cross entropy loss
│  ├── masked_kl_divergence.py  # Masked KL-divergence loss
│  ├── model_util.py            # Some helper functions of model
│  ├── layers/                  # The unique domain data split
│  │   ├── sclstm.py            # SCLSTM network
│  │   ├── encoder.py           # Encoder for SCVAE
│  │   └── decoder_deep.py      # SCLSTM based decoder
├── loader                      # Dataset loader
│  ├── dataset_woz3.py          # Loader of original dataset
│  └── task.py                  # Loader of task and implementation of examplars
├── resource/                   # Data source
│  ├── data_split/*.json        # Data split file
│  ├── feat_*.json              # Feature file containing dialogue acts and slot-values
│  ├── text_*.txt               # Text file with original and delexicalized utterances
│  ├── vacab.txt                # Vocabulary
│  └── template.txt             # The template containing possible domain-action-slot-value
├── ewc.py                      # Compute ewc penalty
├── bleu.py                     # Compute bleu score
├── run_woz3.py                 # Code for starting training, evaluating and testing
├── run.sh                      # Script to run experiment
├── config/*.cfg                # Experiment configuration 
├── experiments/experiment_name # Model and results of experiments
│  ├── model/                   # Models throughout the training of task sequence, notice that the intermediate model.pt contains the test results on all tasks till current task
│  ├── examplars/               # Examplars of each task
│  ├── experiment_name.log      # Training log
│  └── experiment_name.res      # Testing result
├── view_results.ipynb          # Summarize statistics of experiments
└── 
...

Experiment helper functions

.
├── split_multi_task.py                 # Script for generating utterances with unique task (domain/dialogue act), run this only if you need to re-generate feat.json, text.json and data split
├── construct_examplar.py               # Construct exemplars with `herding`, `random` or `prioritized (loss)` based method
└── 
...

Run the code

1. The default datasets have already been available in resource. However, to generate dataset from scratch, please run bash preprocess.sh ${split_type} to generate single domain text, feature and datasplit json file. The split_type can be either do for unique-domain utterances or da for unique dialogue-act utterances.
2. Create your configuration file, the default is config/config.cfg. Please go to config/ for detailed explanations.
3. Specify the hyper-parameters in run.sh. Default settings are provided.
4. Run bash run.sh ${train/test/recover} ${config_file} to reproduce our results or conduct extensive experiments.

Check the result

• Continual learning results in each task are stored in experiments/${experiment_name}/*.log.

• To check aggregated results in our paper, run the notebook view_results.ipynb

• The generated utterances after running test mode are stored in experiments/${experiment_name}/*.res.