Interactive Feature Space Construction using Semantic Information

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 66–74, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Interactive Feature Space Construction using Semantic Information Dan Roth and Kevin Small Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801 {danr,ksmall}@illinois.edu Abstract Specifying an appropriate feature space is an important aspect of achieving good perfor- mance when designing systems based upon learned classifiers. Effectively incorporat- ing information regarding semantically related words into the feature space is known to pro- duce robust, accurate classifiers and is one ap- parent motivation for efforts to automatically generate such resources. However, naive in- corporation of this semantic information may result in poor performance due to increased ambiguity. To overcome this limitation, we introduce the interactive feature space con- struction protocol, where the learner identi- fies inadequate regions of the feature space and in coordination with a domain expert adds descriptiveness through existing semantic re- sources. We demonstrate effectiveness on an entity and relation extraction system includ- ing both performance improvements and ro- bustness to reductions in annotated data. 1 Introduction An important natural language processing (NLP) task is the design of learning systems which per- form well over a wide range of domains with limited training data. While the NLP community has a long tradition of incorporating linguistic information into statistical systems, machine learning approaches to these problems often emphasize learning sophisti- cated models over simple, mostly lexical, features. This trend is not surprising as a primary motivation for machine learning solutions is to reduce the man- ual effort required to achieve state of the art perfor- mance. However, one notable advantage of discrimi- native classifiers is the capacity to encode arbitrarily complex features, which partially accounts for their popularity. While this flexibility is powerful, it often overwhelms the system designer causing them to re- sort to simple features. This work presents a method to partially automate feature engineering through an interactive learning protocol. While it is widely accepted that classifier perfor- mance is predicated on feature engineering, design- ing good features requires significant effort. One un- derutilized resource for descriptive features are ex- isting semantically related word lists (SRWLs), gen- erated both manually (Fellbaum, 1998) and automat- ically (Pantel and Lin, 2002). Consider the follow- ing named entity recognition (NER) example: His father was rushed to [Westlake Hospital]ORG, an arm of [Resurrection Health Care]ORG, in west suburban [Chicagoland]LOC. For such tasks, it is helpful to know that west is a member of the SRWL [Compass Direction] and other such designations. If extracting features using this information, we would require observing only a subset of the SRWL in the data to learn the cor- responding parameter. This statement suggests that one method for learning robust classifiers is to in- corporate semantic information through features ex- tracted from the more descriptive representation: His father was rushed to Westlake [Health Care Institution], an [Subsidiary] of Resur- rection Health Care, [Locative Preposition] [Compass Direction] suburban Chicagoland. 66 Deriving discriminative features from …

Content truncated to 500 words