Automatically Acquiring Phrase Structure Using Distributional Analysis

AUTOMATICALLY ACQUIRING PHRASE STRUCTURE USING DISTRIBUTIONAL ANALYSIS Eric Brill and Mitchell Marcus* Department of Computer Science University of Pennsylvania Philadelphia, Pa. 19104 brill~unagi.cis.upenn.edu, mit ch@unagi.cis.upenn.edu ABSTRACT In this paper, we present evidence that the acquisition of the phrase structure of a natural language is possible without su- pervision and with a very small initial grammar. We describe a language learner that extracts distributional information from a corpus annotated with parts of speech and is able to use this extracted information to accurately parse short sen- tences. The phrase structure learner is part of an ongoing project to determine just how much knowledge of language can be learned solely through distributional analysis. 1. INTRODUCTION This paper is an exploration into the possibility of auto- matically acquiring the phrase structure of a language. We use distributional analysis techniques similar to the techniques originally proposed by Zellig Harris [5] for structural linguists to use as an aid in uncovering the structure of a language. Harris intended his techniques to be carried out by linguists doing field work, as a substitute for what he perceived as unscientific informa- tion gathering by linguists at the time. The procedures Harris describes are intended to uncover "regularities [...] in the distributional relations among the features of speech in question" (page 5). To use distributional analysis to determine empirically whether boy and girl are in the same word class, the linguist would need to determine whether the two words are licensed to occur in the same environments. Harris presented algorithms linguists could use to detect distributionally similar en- tities. Harris did not intend the procedures he proposed to be used as a model of child language acquisition or as a tool for computerized language learning. This would not be feasible because the method Harris describes for deter- mining distributional similarity does not seem amenable to unsupervised acquisition. One way of determining whether boy and girl are in the same word class is to see whether it is the case that for all sentences that boy oc- curs in, the same sentence with girl substituted for boy is an allowable sentence. To do this automatically from *This work was supported by DARPA and AFOSR jointly un- der grant No. AFOSR-90-0066, and by ARO grant No. DAAL 03-89-C0031 PRI. text, one would need a prohibitively large corpus. This lack of sufficient data does not arise in field work be- cause the linguist has access to informants, who are in effect infinite corpora. If one hears the boy finished the homework, the informant can be queried whether the girl finished the homework is also permissible. The procedures Harris outlines for the linguist to use to discover linguistic structure could be used to automat- ically acquire grammatical information if it were possi- ble to do away with the need for a human informant. It is possible that a variation of these procedures could extract information by observing distributional similari- ties in a sufficiently large corpus of unparsed text. In an earlier paper …

Content truncated to 500 words