FactGen: Faithful Text Generation by Factuality-Aware Pre-Training and Contrastive Ranking Fine-Tuning
Abstract
Conditional text generation is supposed to generate a fluent and coherent target text that is faithful to the source text. Although pre-trained models have achieved promising results, they still suffer from the crucial factuality problem. To deal with this issue, we propose a factuality-aware pretraining-finetuning framework named FactGen, which fully considers factuality during two training stages. Specifically, at the pre-training stage, we utilize a natural language inference model to construct target texts that are entailed by the source texts, resulting in a more factually consistent pre-training objective. Then, during the fine-tuning stage, we further introduce a contrastive ranking loss to encourage the model to generate factually consistent text with higher probability. Extensive experiments on three conditional text generation tasks demonstrate the effectiveness and generality of our training framework.
Cite
Text
Lan et al. "FactGen: Faithful Text Generation by Factuality-Aware Pre-Training and Contrastive Ranking Fine-Tuning." Journal of Artificial Intelligence Research, 2023. doi:10.1613/JAIR.1.14267Markdown
[Lan et al. "FactGen: Faithful Text Generation by Factuality-Aware Pre-Training and Contrastive Ranking Fine-Tuning." Journal of Artificial Intelligence Research, 2023.](https://mlanthology.org/jair/2023/lan2023jair-factgen/) doi:10.1613/JAIR.1.14267BibTeX
@article{lan2023jair-factgen,
title = {{FactGen: Faithful Text Generation by Factuality-Aware Pre-Training and Contrastive Ranking Fine-Tuning}},
author = {Lan, Zhibin and Li, Wei and Su, Jinsong and Xiao, Xinyan and Liu, Jiachen and Wu, Wenhao and Lyu, Yajuan},
journal = {Journal of Artificial Intelligence Research},
year = {2023},
pages = {1281-1303},
doi = {10.1613/JAIR.1.14267},
volume = {76},
url = {https://mlanthology.org/jair/2023/lan2023jair-factgen/}
}