Knowledge is Power: Integrating Domain Knowledge and Machine Learning for Robust Textual Inference
Abstract
There is broad recognition of the promise of data-driven and ~intelligent~ solutions to a range of complex decision problems in the physical and life sciences, in engineering, and in their applications to problems in government, defense, and the social sciences. There is agreement that decisions should be informed by data. Indeed, the success of statistical machine learning over thelast few years in several visible supervised function estimation tasks such as image recognition and speech recognition has tilted the pendulum in the direction of ~only data~. Given a task, the standard machine learning methodology suggests collecting annotated data for this task, and thentraining a model for it. However, this methodology is not scalable ~ it is unrealistic to assume it possible to supervise the complex decisions needed without incorporating domain-specific knowledge and background knowledge, which are essential to support problem decomposition, generalization, and transfer, as well as decision making in complex scenarios that have not beenobserved before. Our primary scientific goal is to change this narrow problem setting and offer an encompassing approach to Learning to Reason: supporting decisions made by incorporating knowledge ~domain-specific, commonsense, declarative and statistical ~ along with statistically learned models. We propose to study three fundamental and complementary themes: Inference: We propose a principled computational framework for integrating domain knowledgeand machine learning for supporting diverse complex inferences that are robust and transferable. In particular, our inference framework allows the incorporation of different knowledge representations ~ rules, constraints, and preferences ~ encoding various forms of knowledge ~ quantitative, temporal, relational, and structural. It supports integrating these knowledge representations directly with statistical machine learning models, as a way to bias and combine predictions of individual models, or within a joint (end-to-end) learning and inference framework that accounts for the declarative constraints while learning.Knowledge Representation: Declarative knowledge representations are powerful since they often come with a proof system. However, they do not account for the ambiguity and variability that are inherent in the messy data that must be dealt with in real world applications. We propose to develop a knowledge representation scheme that supports the ~traditional~ relational structure, but nevertheless enjoys useful properties of continuous representations. Learning with Knowledge is perhaps the most significant challenge to effective machine learning for complex problems. We will address two key issues in this context: (i) How to guide learning of complex models, via (declarative, soft, and hard) prior knowledge and indirect supervision, so that interdependent output variables satisfy, perhaps softly, global expectations;and (ii) How to decompose models and automatically learn KR-like structured models of knowledge, in the course of performing a complex task, as a way to better support transfer across domains and facilitate better generalization for decisions in previously unseen situations. The PI is a leading researcher in Artificial Intelligence, with expertise and major contributions in natural language processing, machine learning, and knowledge representation and reasoning.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2019
- Source ID
- N000141912620
Entities
People
- Dan Roth
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania