Partially-Aligned Data-to-Text Generation with Distant Supervision

[PDF] [Video] [Slides]

This is the code for the EMNLP 2020 paper "Partially-Aligned Data-to-Text Generation with Distant Supervision". Traditional text generation task requires well-aligned data which is expensive to annotate. We relax the strict restrictions and propose this new task aiming at utilizing automatically made partially-aligned data. This method considerably expands the application domains where only automatically partially-aligned data is available.

Requirements

• GCC >= 4.8
• Python >= 3.7

Install

git clone https://github.com/fuzihaofzh/distant_supervision_nlg.git
cd distant_supervision_nlg
./scripts/setup.sh

Preprocess Data

./scripts/preprocess.sh wita50k

Train Baseline Model (Optional)

The model will be evaluated automatically during training.

# Train S2ST model
./scripts/train.sh wita50k base
# Check Score
tail -n 1 output/eval/wita50k__base/eval.100.txt

Train Our DSG Model

The model will be evaluated automatically during training.

# Step 1. SE Training
./scripts/train.sh wita50k endorsement,pretrain
# Step 2. S2SG Training
./scripts/train.sh wita50k endorsement,beam_endorse
# Check Score
tail -n 1 output/eval/wita50k__endorsement,beam_endorse/eval.100.txt

Cite

@inproceedings{fu2020partially,
  title={Partially-Aligned Data-to-Text Generation with Distant Supervision},
  author={Fu, Zihao and Shi, Bei and Lam, Wai and Bing, Lidong and Liu, Zhiyuan},
  booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  pages={9183--9193},
  year={2020}
}