Machine Learning Thesis Proposal
Speaker
HAITIAN SUN
Ph.D. Student
Machine Learning Department
Carnegie Mellon University
When
-
Where
Virtual Presentation - ET
Description
Answering complex questions posed in natural language requires understanding semantic information in questions and contexts and perform various reasoning steps to find answers. In this proposal, we discuss several methods for challenges in building intelligent Question Answering (QA) systems to answer complex questions. To begin with, we study the task of logical reasoning over symbolic knowledge bases (KBs). We first propose a neural reasoning model for QA in symbolic space, and then discuss possibilities to extend it to embedding space to allow more relaxation and generalization. Since symbolic KBs are commonly incomplete, we propose another method to construct virtual KBs (VKB) from text that support most reasoning tasks as regular KBs. We additionally propose methods to integrate reasoning modules into language models (LMs) for more challenging QA tasks. The reasoning tasks discussed previously, however, mainly focus on QA with complete information, i.e. questions are well-defined such that deterministic answers exist. In the final part of this proposal, we consider a more challenging task where questions are asked with incomplete information. We release a new dataset for this task and show the challenges faced by current QA systems. We propose models to tackle two main challenges in this task: (1) answering questions with unsatisfied constraints, and (2) distinguishing context-dependent answers.
Thesis Committee:
Ruslan Salakhutdinov (Chair)
William W. Cohen
Emma Strubell
Hannaneh Hajishirzi (University of Washington)
Zoom Participation. See announcement.