Bracketing Input for Accurate Parsing

Bracketing Input for Accurate Parsing * Yongkyoon No a Chungnam National University 220 Koong-dong, Yuseong-gu, Taejeon, Korea yno@linguist.cnu.ac.kr Abstract. Syntax parsers can benefit from speakers’ intuition about constituent structures indicated in the input string in the form of parentheses. Focusing on languages like Korean, whose orthographic convention requires more than one word to be written without spaces, we describe an algorithm for passing the bracketing information across the tagger to the probabilistic CFG parser, together with one for heightening (or penalizing, as the case may be) probabilities of putative constituents as they are suggested by the parser. It is shown that two or three constituents marked in the input suffice to guide the parser to the correct parse as the most likely one, even with sentences that are considered long. Keywords: manually parsed corpus, Probabilistic Context Free Grammar, Korean syntax, bottom-up chart parser, pre-annotated input, Paak, KWGInterpreter 1. Introduction Syntactic parsers often produce parses many of whose parts are correct but which do contain partially inaccurate structures. To attain a high level of precision and recall, serious research has to be conducted and just about every kind of research in this field may well benefit from a moderate amount of manually parsed, i.e, correctly parsed, sentences. When it comes to the Korean language, a widely available, yet dependable corpus of parsed sentences is something that has yet to be constructed. In this paper, we describe a move that will help build a parsed corpus of about one thousand sentences. Our idea here is to use an initial stage probabilitic CFG parser with a partially bracketed input. This work is inspired by Pereira and Schabes (1992), who use bracketed sentences to induce grammar rules automatically. We will describe our initial stage parser, Paak, in Section 2. How the input string with some bracketings are processed in this parser is described in the next section. Section 4 gives algorithms for keeping track of the bracketing information and for affecting the probabilities of constituents, which conform to, are consistent with, or contradict the information annotated in the input. The performance gains at each degree of bracketing information are estimated in Section 5, where it is shown that no more than two pairs of bracketings suffice to yield the correct parse of a sentence of length twenty. 2. The probabilistic CFG parser Paak Paak is a probabilistic parser of the Korean language with a set of context-free grammar rules. It relies on a tagger of Korean for its analysis of input string into the smallest units of syntactic * I am grateful to one of the anonymous referees of the Program Committee of PACLIC 21, who gave me useful comments. Copyright 2007 by Yongkyoon No 358 processing.1 The tagger uses fifty three tags, many of which correspond to parts-of-speech. Paak has 45 preterminal symbols and 20 nonterminals. Currently it has just over three hundred PCFG rules. The parser is a probabilistic parser implemented in Java. The implementation basically follows Charniak(1993: 1.4). It is a …

Content truncated to 500 words