Unified Structure Generation for Universal Information Extraction

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5755 - 5772 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Unified Structure Generation for Universal Information Extraction Yaojie Lu1,4,∗, Qing Liu1,4,∗, Dai Dai3, Xinyan Xiao3, Hongyu Lin1,†, Xianpei Han1,2,5, Le Sun1,2,†, Hua Wu3 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3Baidu Inc., Beijing, China 4University of Chinese Academy of Sciences, Beijing, China 5Beijing Academy of Artificial Intelligence, Beijing, China {yaojie2017,liuqing2020,hongyu,xianpei,sunle}@iscas.ac.cn {daidai,xiaoxinyan,wu_hua}@baidu.com Abstract Information extraction suffers from its varying targets, heterogeneous structures, and demand- specific schemas. In this paper, we propose a unified text-to-structure generation framework, namely UIE, which can universally model dif- ferent IE tasks, adaptively generate targeted structures, and collaboratively learn general IE abilities from different knowledge sources. Specifically, UIE uniformly encodes different extraction structures via a structured extraction language, adaptively generates target extrac- tions via a schema-based prompt mechanism – structural schema instructor, and captures the common IE abilities via a large-scale pre- trained text-to-structure model. Experiments show that UIE achieved the state-of-the-art per- formance on 4 IE tasks, 13 datasets, and on all supervised, low-resource, and few-shot set- tings for a wide range of entity, relation, event and sentiment extraction tasks and their unifi- cation. These results verified the effectiveness, universality, and transferability of UIE1. 1 Introduction Information extraction (IE) aims to identify and structure user-specified information from unstruc- tured texts (Andersen et al., 1992; Grishman, 2019). IE tasks are highly diversified due to its varying targets (entity, relation, event, sentiment, etc.), het- erogeneous structures (spans, triplets, records, etc.), and demand-specific schemas (Grishman and Sund- heim, 1996; Mitchell et al., 2005; Ji and Grishman, 2011). Currently, most IE approaches are task- specialized, which leads to dedicated architec- tures, isolated models, and specialized knowl- ∗ Part of this work was done when Yaojie Lu and Qing Liu interned at Baidu. †Corresponding authors. 1https://universal-ie.github.io Entity PER: Steve ORG: Apple Relation (Work for, Steve, Apple) Event (Start-Position: become …） Sentiment Positive: excited Text Unified Generation Task Schema Instance Entity PER: _ ORG:_ Relation (_, Work for, _) Event Sentiment Positive { Opinion: _; Target: _ } In 1997, Steve was excited to become the CEO of Apple. In 1997, Steve was excited to become the CEO of Apple. Type Start Position employee employer … Positive Opinion Target (a) Task-specialized IE (b) Universal IE Work For In 1997, Steve was excited to become the CEO of Apple. Start-Position PER ORG In 1997, Steve was excited to become the CEO of Apple. Entity Person Figure 1: From (a) Task-specialized IE: different tasks, different structures, different schemas to (b) Universal IE: unified modeling via structure generation. edge sources for different IE task. These task- specialized solutions greatly hinder the rapid ar- chitecture development, effective knowledge shar- ing, and quick cross-domain adaptation of IE sys- tems. First, it is very complicated to develop dedicated architectures for a large amount of IE tasks/settings/scenarios. Second, learning …

Content truncated to 500 words