A Plethora of Methods for Learning English Countability

A Plethora of Methods for Learning English Countability Timothy Baldwin CSLI Stanford University Stanford, CA 94305 USA tbaldwin@csli.stanford.edu Francis Bond NTT Communication Science Laboratories Nippon Telegraph and Telephone Corporation Kyoto, Japan bond@cslab.kecl.ntt.co.jp Abstract This paper compares a range of methods for classifying words based on linguis- tic diagnostics, focusing on the task of learning countabilities for English nouns. We propose two basic approaches to feature representation: distribution-based representation, which simply looks at the distribution of features in the cor- pus data, and agreement-based represen- tation which analyses the level of token- wise agreement between multiple pre- processor systems. We additionally com- pare a single multiclass classifier archi- tecture with a suite of binary classifiers, and combine analyses from multiple pre- processors. Finally, we present and evalu- ate a feature selection method. 1 Introduction Lexical acquisition can be described as the process of populating a grammar skeleton with lexical items, through a process of mapping word lemmata onto lexical types described in the grammar. Depending on the linguistic precision of the base grammar, lex- ical acquisition can range in complexity from sim- ple part-of-speech tagging (shallow lexical acquisi- tion) to the acquisition of selectionally-constrained subcategorisation frame clusters or constructional compatibilities (deep lexical acquisition). Our par- ticular interest is in the latter task of deep lexical acquisition with respect to English nouns. We are interested in developing learning tech- niques for deep lexical acquisition which take a fixed set of linguistic diagnostics, and classify words ac- cording to corpus data. We propose a range of gen- eral techniques for this task, as exemplified over the task of English countability acquisition. Countabil- ity is the syntactic property that determines whether a noun can take singular and plural forms, and af- fects the range of permissible modifiers. Many nouns have both countable and uncountable lemmas, with differences in meaning: I submitted two papers “documents” (countable) vs. Please use white paper “substance to be written on” (uncountable). This research complements that described in Baldwin and Bond (2003), where we present the lin- guistic foundations and features drawn upon in the countability classification task, and motivate the claim that countability preferences can be learned from corpus evidence. In this paper, we focus on the methods used to tackle the task of countability classification based on this fixed feature set. The remainder of this paper is structured as fol- lows. Section 2 outlines the countability classes, resources and pre-processors. Section 3 presents two methods of representing the feature space. Sec- tion 4 details the different classifier designs and the dataset, which are then evaluated in Section 5. Fi- nally, we conclude the paper with a discussion in Section 6. 2 Preliminaries In this section, we describe the countability classes, the resources used in this research, and the feature extraction method. These are described in greater detail in Baldwin and Bond (2003). 2.1 Countability classes Nouns are classified as belonging to one or more of four possible classes: countable, uncountable, plural only and bipartite. Countable nouns can be …

Content truncated to 500 words