Bilingual Parsing with Factored Estimation: Using English to Parse Korean

Bilingual Parsing with Factored Estimation: Using English to Parse Korean David A. Smith and Noah A. Smith Department of Computer Science Center for Language and Speech Processing Johns Hopkins University Baltimore, MD 21218, USA {d,n}asmith@cs.jhu.edu Abstract We describe how simple, commonly understood statisti- cal models, such as statistical dependency parsers, proba- bilistic context-free grammars, and word-to-word trans- lation models, can be effectively combined into a uni- fied bilingual parser that jointly searches for the best En- glish parse, Korean parse, and word alignment, where these hidden structures all constrain each other. The model used for parsing is completely factored into the two parsers and the TM, allowing separate parameter es- timation. We evaluate our bilingual parser on the Penn Korean Treebank and against several baseline systems and show improvements parsing Korean with very lim- ited labeled data. 1 Introduction Consider the problem of parsing a language L for which annotated resources like treebanks are scarce. Suppose we have a small amount of text data with syntactic an- notations and a fairly large corpus of parallel text, for which the other language (e.g., English) is not resource- impoverished. How might we exploit English parsers to improve syntactic analysis tools for this language? One idea (Yarowsky and Ngai, 2001; Hwa et al., 2002) is to project English analysis onto L data, “through” word-aligned parallel text. To do this, we might use an English parser to analyze the English side of the parallel text and a word-alignment algorithm to induce word cor- respondences. By positing a coupling of English syntax with L syntax, we can induce structure on the L side of the parallel text that is in some sense isomorphic to the English parse. We might take the projection idea a step farther. A statistical English parser can tell us much more than the hypothesized best parse. It can be used to find every parse admitted by a grammar, and also scores of those parses. Similarly, translation models, which yield word alignments, can be used in principle to score competing alignments and offer alternatives to a single-best align- ment. It might also be beneficial to include the predic- tions of an L parser, trained on any available annotated L data, however few. This paper describes how simple, commonly under- stood statistical models—statistical dependency parsers, probabilistic context-free grammars (PCFGs), and word translation models (TMs)—can be effectively combined into a unified framework that jointly searches for the best English parse, L parse, and word alignment, where these hidden structures are all constrained to be consistent. This inference task is carried out by a bilingual parser. At present, the model used for parsing is completely fac- tored into the two parsers and the TM, allowing separate parameter estimation. First, we discuss bilingual parsing (§2) and show how it can solve the problem of joint English-parse, L-parse, and word-alignment inference. In §3 we describe param- eter estimation for each of the factored models, includ- ing novel applications of log-linear models to English dependency parsing and Korean …

Content truncated to 500 words