Class-Driven Attribute Extraction

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 921–928 Manchester, August 2008 Class-Driven Attribute Extraction Benjamin Van Durme, Ting Qian and Lenhart Schubert Department of Computer Science University of Rochester Rochester, NY 14627, USA Abstract We report on the large-scale acquisition of class attributes with and without the use of lists of representative instances, as well as the discovery of unary attributes, such as typically expressed in English through prenominal adjectival modifica- tion. Our method employs a system based on compositional language processing, as applied to the British National Corpus. Ex- perimental results suggest that document- based, open class attribute extraction can produce results of comparable quality as those obtained using web query logs, indi- cating the utility of exploiting explicit oc- currences of class labels in text. 1 Introduction Recent work on the task of acquiring attributes for concept classes has focused on the use of pre- compiled lists of class representative instances, where attributes recognized as applying to multi- ple instances of the same class are inferred as be- ing likely to apply to most, or all, members of that class. For example, the class US President might be represented as a list containing the en- tries Bill Clinton, George Bush, Jimmy Carter, etc. Phrases such as Bill Clinton’s chief of staff ..., or search queries such as chief of staff bush, provide evidence that the class US President has as an at- tribute chief of staff. Usually the focus of such systems has been on on binary attributes, such as the example chief of staff, while less attention has been paid to unary c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. class attributes such as illegal for the class Drug, or warm-blooded for the class Animal.1 These attributes are most typically expressed in English through prenominal adjectival modification, with the nominal serving as a class designator. When attribute extraction is based entirely on instances and not the class labels themselves, this form of modification goes undiscovered. In what follows we explore both the impact of gazetteers in attribute extraction as well as the acquisition and filtering of unary class attributes, through a process based on logical form genera- tion from syntactic parses derived from the British National Corpus. 2 Extraction Framework Extraction was performed using a modified ver- sion of the KNEXT system, a knowledge acquisi- tion framework constructed for large scale genera- tion of abstracted logical forms through composi- tional linguistic analysis. The following provides an overview of KNEXT and its target knowledge representation, Episodic Logic. 2.1 Episodic Logic Automatically acquiring general world knowledge from text is not a task that provides an immedi- ate solution to any real world problem.2 Rather, the motivation for acquiring large stores of back- ground knowledge is to enable research within other areas of artificial intelligence, e.g., the con- struction of systems that can engage in dialogues about everyday topics in unrestricted English, use 1Almuhareb and Poesio …

Content truncated to 500 words