Back to search

Cornell University

Structured State Tracking for Natural Language Understanding

Abstract

dc:description.abstract

Autonomous agents that collaborate with humans must understand language, track the state of the world, and make good decisions. A central challenge common to these three desiderata is uncertainty. Language is often ambiguous, with many possible interpretations of the same utterance. The state of the world is unobservable, as a single agent cannot observe every aspect due to physical constraints. Agents must reason about the unobserved aspects as new information is received. Finally, many decisions have uncertain outcomes. Agents must anticipate how the world will respond to their decisions in order to achieve the best outcomes. In this thesis, we build uncertainty-aware agents that understand language, the state of the world, and decision-making.The thesis is divided into three sections, with each section exploring a different application by applying a modern twist to a classical approach. The first section studies uncertainty-aware representations in language modeling. We revisit classical structured and uncertainty-aware language models, showing that they can achieve strong performance when combined with modern techniques from deep learning. The second section studies the use of language models for reasoning about uncertainty in question-answering. We reason about unobserved reasoning paths, applying modern pretrained language model representations to aid in high-level reasoning. The third section studies task-oriented dialogue as decision-making under uncertainty. We intertwine classical decision-making methods with the powerful language and code understanding capabilities of modern large language models, resulting in an accurate neurosymbolic dialogue system.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Computer Science
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Computer Science
Grantor
Cornell University
Year dc:date.issued
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Chiu, Justin
Chair dc:contributor.chair
  • Rush, Alexander
Committee members dc:contributor.committeemember
  • Lee, Daniel
  • Choudhury, Sanjiban

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
16611792
Dc Identifier Other
ProQuest Submission ID: 14389
ProQuest Publication ID: 31333043

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Chiu, Justin. Structured State Tracking for Natural Language Understanding. Doctor of Philosophy thesis, Cornell University, 2024. https://hdl.handle.net/1813/116417