Spanish HPSG Treebank based on the AnCora Corpus

Spanish HPSG Treebank based on the AnCora Corpus Luis Chiruzzo, Dina Wonsever Universidad de la Rep´ublica Montevideo, Uruguay {luischir, wonsever}@fing.edu.uy Abstract This paper describes a corpus of HPSG annotated trees for Spanish that contains morphosyntactic information, annotations for semantic roles, clitic pronouns and relative clauses. The corpus is based on the Spanish AnCora corpus, which contains trees for 17,000 sentences comprising half a million words, and it has CFG style annotations. The corpus is stored in two different formats: An XML dialect that is the direct serialization of the typed feature structure trees, and an HTML format that is suitable for visualizing the trees in a browser. Keywords: HSPG, Spanish, AnCora, corpus 1. Introduction We present the construction of a Spanish HPSG treebank based on the Spanish corpus AnCora (Taul´e et al., 2008). This is part of an ongoing project to build a statistical Span- ish HPSG parser. An earlier version of this corpus is described in (Chiruzzo and Wonsever, 2016). In this previous work we used a se- ries of hand-crafted rules to process all constituents in An- Cora in order to find the heads and binarize all the phrases, identifying when the constituents acted as specifier, com- plement or modifier. In this first step, clitics and relative clauses had been identified but not properly handled. In this work we finished the transformation of the corpus by incorporating an analysis for clitics and relative clauses into the feature structure and by annotating those constructions in the corpus, as well as producing a way of visualizing the HPSG trees. 2. Background HPSG grammars are rich grammars (Pollard and Sag, 1994) that are able to represent both syntactic and seman- tic information in the same parse tree. The nodes in the trees and the rules to combine them are defined as typed feature structures (Carpenter, 1992) with a unification op- eration. The leaves of the parse tree are the words of a sen- tence, and the HPSG feature structure generally includes morphosyntactic information such as part of speech, agree- ment features, and syntactic valence features. The English Resource Grammar (Flickinger, 1999) is an implementation of the HPSG principles for English, built into the LinGO Grammar Matrix (Bender et al., 2002), a framework for building unification grammars. There also exists a Spanish HPSG grammar built over the same prin- ciples: the Spanish Resource Grammar (SRG) (Marimon, 2010a). Both grammars are hand crafted paying particular attention to the linguistic details of the theory and the cor- rectness of the modeled sentences. This means the trees ob- tained using these grammars are very rich, but on the other hand sentences that are not perfectly written (as is the case for general text extracted from the web) would not be even partially parsed. In this work we have the final aim of build- ing a statistical parser from scratch based on the statistics of a properly annotated corpus, so we aim to improve the robustness of the parsing process for sentences that not nec- essarily …

Content truncated to 500 words