General Collaborative Framework between Large Language Model and Experts for Universal Information Extraction

Anonymous

Abstract: Recently, unified information extraction have been widely concerned NLP community, which aims at using a unified paradigm to perform various information extraction tasks. However, they inevitably suffering from some thorny problems such as noise interference, abstract label semantics, and diverse span granularity. In this paper, First of all, we start by presenting three problematic assumptions that exist in previous research works from a unified information extraction perspective. These problems severely hinder the development of information extraction models. Furthermore, to solve these problems, we propose the General Collaborative Information Extraction framework for universal information extraction. Specifically, GCIE consists of a general Recognizer for identifying predefined types and multiple task-specific Experts for extracting spans. The Recognizer is a large language model, while the Expert is a series of smaller language models, and they collaborate in a pipeline to achieve unified or task-specific information extraction. Empirical experiments on 6 IE tasks and 13 datasets, under supervised and few-shot settings, validate the effectiveness and generality of our approach.

Paper Type: long

Research Area: Information Extraction

Contribution Types: Model analysis & interpretability

Languages Studied: English