Reading to Learn: Constructing Features from Semantic Abstracts

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 958–967, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Reading to Learn: Constructing Features from Semantic Abstracts Jacob Eisenstein∗ James Clarke† Dan Goldwasser† Dan Roth∗† ∗Beckman Institute for Advanced Science and Technology, †Department of Computer Science University of Illinois Urbana, IL 61801 {jacobe,clarkeje,goldwas1,danr}@illinois.edu Abstract Machine learning offers a range of tools for training systems from data, but these methods are only as good as the underly- ing representation. This paper proposes to acquire representations for machine learn- ing by reading text written to accommo- date human learning. We propose a novel form of semantic analysis called read- ing to learn, where the goal is to obtain a high-level semantic abstract of multi- ple documents in a representation that fa- cilitates learning. We obtain this abstract through a generative model that requires no labeled data, instead leveraging repe- tition across multiple documents. The se- mantic abstract is converted into a trans- formed feature space for learning, result- ing in improved generalization on a rela- tional learning task. 1 Introduction Machine learning offers a range of powerful tools for training systems to act in complex environ- ments, but these methods depend on a well-chosen representation for features. For learning to suc- ceed the representation often must be crafted with knowledge about the application domain. This poses a bottleneck, requiring expertise in both ma- chine learning and the application domain. How- ever, domain experts often express their knowl- edge through text; one direct expression is through text designed to aid human learning. In this paper we exploit text written by domain experts in or- der to build a more expressive representation for learning. We term this approach reading to learn. The following scenario demonstrates the moti- vation for reading to learn. Imagine an agent given a task within its world/environment. The agent has no prior knowledge of the task but can perceive the world through low-level sensors. Learning directly from the sensors may be difficult, as interesting tasks typically require a complex combination of sensors. Our goal is to acquire domain knowledge through the semantic analysis of text, so as to pro- duce higher-level relations through combinations of sensors. As a concrete example consider the problem of learning how to make legal moves in Freecell soli- taire. Relevant sensors may indicate if an object is a card or a freecell, whether a card is a certain value, and whether two values are in sequence. Although it is possible to express the rules with a combination of sensors, learning this combina- tion is difficult. Text can facilitate learning by pro- viding relations at the appropriate level of gen- eralization. For example, the sentence: “You can place a card on an empty freecell,” suggests not only which sensors are useful together but also how these sensors should be linked. Assuming the sensors are represented as predicates, one possi- ble relation this sentence suggests is: r(x, y) = card(x) ∧ freecell(y) ∧ empty(y). …

Content truncated to 500 words