A Treebank of Spanish and its Application to Parsing

This paper presents joint research between a Spanish team and an American one on the development and exploitation of a Spanish treebank. Such treebanks for other languages have proven valuable for the development of high-quality parsers and for a wide variety of language studies. However, when the project started, at the end of 1997, there was no syntactically annotated corpus for Spanish. This paper describes the design of such a treebank and its initial application to parser construction. 1. Constructing a Spanish treebank 1.1. Preliminary considerations This paper presents joint research between a Spanish team and an American one on the development and exploitation of a Spanish treebank. Such treebanks for other languages have proven valuable for the development of high-quality parsers and for a wide variety of language studies. As there was no previous experience in building a syntactically annotated corpus for Spanish, the first effort consisted necessarily in writing a set of annotation guidelines. The starting point was the existing documentation at that time, especially the Penn Treebank project (Marcus, Santorini and Marcinkiewicz, 1993; Bies et al., 1995), the EAGLES preliminary recommendations (EAGLES, 1996), and the Negra corpus (Skut et al., 1997). Our experience in developing Spanish NLP systems told us that a pure phrase structure annotation (typical of the English treebanks) would not be enough for inducing relevant rules for Spanish. At the least, information about agreement and syntactic functions is necessary for Spanish, and we wanted to incorporate that information in our trees in the form of features. The treebank has been created mostly by hand, although some automatic pre-tagging of the data is performed, as described below, to speed treebank creation.