Plan-then-Generate: Controlled Data-to-Text Generation via Planning
Authors: Yixuan Su, David Vandyke, Sihui Wang, Yimai Fang, and Nigel Collier
Code for EMNLP 2021 paper Plan-then-Generate: Controlled Data-to-Text Generation via Planning
1. Environment Setup:
(1) Hardware Requirement:
The code in this repo is thoroughly tested on our machine with a single Nvida V100 GPU (16GB)
(2) Installation:
chmod +x ./config_setup.sh
./config_setup.sh2. ToTTo Data Preprocessing:
Option (1): Preprocess the ToTTo data from scratch by yourself:
cd ./data
chmod +x ./prepare_data.sh
./prepare_data.shThis process could take up to 1 hour
Option (2): Download the our processed data here
unzip data.zip and replace with the empty ./data folderFor more details about ToTTo dataset, please refer to the original Google Research repo
3. Content Planner:
Please refer to README.md in ./content_planner folder
4. Sequence Generator:
Please refer to README.md in ./generator folder
5. Citation
If you find our paper and resources useful, please kindly cite our paper:
@inproceedings{su-etal-2021-plan-generate,
title = "Plan-then-Generate: Controlled Data-to-Text Generation via Planning",
author = "Su, Yixuan and
Vandyke, David and
Wang, Sihui and
Fang, Yimai and
Collier, Nigel",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
month = nov,
year = "2021",
address = "Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.findings-emnlp.76",
pages = "895--909",
abstract = "Recent developments in neural networks have led to the advance in data-to-text generation. However, the lack of ability of neural models to control the structure of generated output can be limiting in certain real-world applications. In this study, we propose a novel Plan-then-Generate (PlanGen) framework to improve the controllability of neural data-to-text models. Extensive experiments and analyses are conducted on two benchmark datasets, ToTTo and WebNLG. The results show that our model is able to control both the intra-sentence and inter-sentence structure of the generated output. Furthermore, empirical comparisons against previous state-of-the-art methods show that our model improves the generation quality as well as the output diversity as judged by human and automatic evaluations.",
}