Genie: Achieving Human Parity in Content-Grounded Datasets Generation
This addresses a major obstacle for advancing content-grounded generation tasks by providing scalable, high-quality synthetic data, though it is incremental as it builds on existing data generation methods.
The paper tackles the lack of high-quality data for content-grounded generation tasks by proposing Genie, a method for automatically generating synthetic data, and shows that models trained on this data achieve parity with or outperform those trained on human-written data in tasks like long-form question-answering and summarization, with consistent gains in faithfulness.
The lack of high-quality data for content-grounded generation tasks has been identified as a major obstacle to advancing these tasks. To address this gap, we propose Genie, a novel method for automatically generating high-quality content-grounded data. It consists of three stages: (a) Content Preparation, (b) Generation: creating task-specific examples from the content (e.g., question-answer pairs or summaries). (c) Filtering mechanism aiming to ensure the quality and faithfulness of the generated data. We showcase this methodology by generating three large-scale synthetic data, making wishes, for Long-Form Question-Answering (LFQA), summarization, and information extraction. In a human evaluation, our generated data was found to be natural and of high quality. Furthermore, we compare models trained on our data with models trained on human-written data -- ELI5 and ASQA for LFQA and CNN-DailyMail for Summarization. We show that our models are on par with or outperforming models trained on human-generated data and consistently outperforming them in faithfulness. Finally, we applied our method to create LFQA data within the medical domain and compared a model trained on it with models trained on other domains.