.nodeKeep

factgen-pub.pdf

Version of Record

External Links

Node Root

Zhibin Lan

Wei Li

Jinsong Su

Xinyan Xiao

Jiachen Liu

Wenhao Wu

Yajuan Lyu

root

Research Report

This scientific article introduces FactGen, a novel factuality-aware pretraining-finetuning framework for conditional text generation. By utilizing a natural language inference model during the pre-training stage and introducing a contrastive ranking loss during the fine-tuning stage, FactGen achieves fluent and coherent text generation that remains faithful to the source material. Extensive experiments on three conditional text generation tasks demonstrate the effectiveness and generality of this training framework.

FactGen