Large language models (LLMs) can perform a new task by merely conditioning on task instructions and a few input-output examples, without optimizing any parameters. This is called In-Context Learning (ICL). In-context Information Extraction has recently garnered attention in the research community. However, current experiment results are generally suboptimal. We attribute this primarily to the fact that the complex task settings and a variety of edge cases are hard to be fully expressed in the length-limited context. In this paper, we propose a Guideline Learning (GL) framework for In-context IE which learns to generate and follow guidelines. During the learning phrase, GL automatically synthesizes a set of guidelines from a few annotations, and during inference, helpful guidelines are retrieved for better ICL.
View on arXiv