Constituent Structure for Filipino: Induction through Probabilistic Approaches

Constituent Structure for Filipino: Induction through Probabilistic Approaches* Danniel Alcantara and Allan Borra College of Computer Studies, De La Salle University, 2401 Taft Avenue, 1004 Manila, Philippines {alcantarad, borraa}@ dlsu.edu.ph Abstract. The current state of Philippine linguistic resources, which includes formal grammars, electronic dictionaries and corpora are not yet significant to address industrial- strength language technologies. This paper discusses a computational approach in automatically estimating constituent structures from a corpus using unsupervised probabilistic approaches. Two models are presented and results show an F1 measure of greater than 69%. Issues and phenomena of the Filipino language are identified and discussed Keywords: Computational Linguistics, Probabilistic Approach, Constituent Structure, Context Free Grammar, Grammar Induction 1. Introduction This paper discusses the algorithms used for the automatic induction of grammar for the Filipino language. The rationale for the study stems on the minimal work done on the development of a computational grammar for the Filipino language for the development of robust and industrial- strength natural language analysis and technologies. Existing Filipino grammars can only handle a subset of declarative type sentences. Considering the difficulty to manually construct a robust grammar capable of parsing a broad scope of sentences, automatic grammar induction is a consideration that can be used for learning language structure. Automated grammar induction systems deals with the generation of a grammar based from input corpora. Existing work for grammar induction fall under two categories based on their input constraints: Supervised and Unsupervised. Slightly supervised systems generate grammar rules from bracketed corpora or tree banks. Bracketed corpora are text documents that have been bracketed by a linguist to represent the skeletal syntactic structure of the sentences. Treebanks are large corpora that have been annotated with the part of speech tags, syntactic structure, and other functional attributes necessary. Unsupervised systems make use of non- bracketed corpora while applying searching and clustering algorithms to attempt to learn the language rules. Works on slightly supervised grammar induction, such as the works of Lari and Young(1991), Brill(1993), Sekine and Grishman(1995), and Charniak(1996), present different output formalisms to represent the grammar. However, Filipino is currently a resource-limited language and does not have the computational resources necessary for the algorithms presented. There are existing corpora available for the language, but these have not yet been bracketed. * The authors wish to thank Dr. Shirley Dita for providing the initial gold standard. Copyright 2008 by Danniel Alcantara and Allan Borra 113 22nd Pacific Asia Conference on Language, Information and Computation, pages 113–122 Osborne and Briscoe(1997), Clark(2001), Klein and Manning,(2001) attempt to learn the grammar formalism for the language through the use of statistical analysis methods applied to a tagged corpus. Klein and Manning(2002) applies a context-constituency model and achieved promising results, precision rate of 55% and recall rate of 48%. Existing systems are designed for the English language and have had modifications applied that are specific to the right-biased nature and Subject-Verb-Object structure. The Filipino language does not follow these structures; rather it has free word order patterns, and Predicate-Topic phenomena, …

Content truncated to 500 words