Back to search

University of Waterloo

Information Extraction for Low-Resource Schemas

Abstract

dc:description.abstract

Information Extraction (IE) is a set of important tasks in the study of creating structured data such as knowledge graphs from unstructured data such as text. The past paradigm of IE focused on models with specialized neural network architectures, usually based on transformer encoders. These models typically focus on a single subtask of IE, following a single schema of entity and relation types, and are trained via supervised learning on large datasets of annotated texts. Meanwhile, the current paradigm of IE, called Universal IE (UIE), involves large language models which can generalize across IE subtasks and to completely unseen schemas, but which lack other abilities such as entity grounding and calibration. We first discuss structural consistency, a new measure of robustness in information extraction based on compositionality. We present structural consistency post-training (SCPT) as a data augmentation method to boost structural consistency for a wide range of model architectures. Besides greatly improving robustness, SCPT significantly reduces the amount of labelled data needed to achieve the same level of performance when training specialized IE models. Second, we use reasoning-based data augmentation techniques to gather AdaIE, a very large collection of human-annotated information extraction schemas. We diverge from UIE and align the dataset with a new task we call Guided Information Extraction (GIE). GIE emphasizes the tight grounding and schema-following requirements which have been largely neglected in UIE. Evaluations of state-of-the-art UIE models reveal that state of the art UIE methods can be surpassed by recent commercial large language models (LLMs). Although those LLMs achieve below human performance on AdaIE, they are rapidly advancing. Overall, we hope that both works presented will steer the IE research community towards unifying the strengths of the old and new IE paradigms, while casting light on their weaknesses.

Degree

thesis:*
Grantor dc:publisher
University of Waterloo
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Xu, Justin

Subjects

dc:subject × 9

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/10012/23544

Chain of custody

source
Harvested from
University of Waterloo
Base URL
dspacemainprd01.lib.uwaterloo.ca/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Xu, Justin. Information Extraction for Low-Resource Schemas. University of Waterloo, 2026. https://hdl.handle.net/10012/23544