A New Version of the Składnica Treebank of Polish Harmonised with the Walenty Valency Dictionary

A New Version of the Składnica Treebank of Polish Harmonised with the Walenty Valency Dictionary Marcin Woliński, Elżbieta Hajnicz, Tomasz Bartosiak Institute of Computer Science, Polish Academy of Sciences Jana Kazimierza 5, 01-248 Warsaw, Poland {wolinski,hajnicz}@ipipan.waw.pl, tomasz.bartosiak@gmail.com Abstract This paper reports on developments in the Składnica treebank of Polish which were possible due to the switch to the Walenty valency dictionary. The change required several modifications in the Świgra parser, such as implementing unlike coordination, semantically motivated phrases, and non-standard case values. A procedure to upgrade manually disambiguated trees of Składnica was required as well. Modifications introduced in the treebank included systematic changes of notation and resolving ambiguity between semantically motivated phrases. The procedure of confronting Składnica treebank with the trees generated with the new version of the Świgra parser using Walenty dictionary allowed us to check the consistency of all the resources. This resulted in several corrections introduced in both the treebank and the valence dictionary. Keywords: treebank, constituency parsing, valency dictionary 1. Introduction This article concerns adapting the Składnica treebank to the valence dictionary Walenty. Składnica predates Walenty, so initially the treebank was based on the valence dictionary of Świdziński (1994). Walenty surpasses this dictionary both in size and the number of linguistic phenomena being represented. Therefore, deploying Walenty was an obvious choice for the further development of Składnica. The procedure to adapt Składnica to the new dictionary was to a large extent automatic. However, the differences be- tween the resources made it necessary to manually correct some parse trees. In this article, we present the method of automatic mapping and the problems that needed manual intervention. The article is organised as follows. First we present the re- sources – the treebank Składnica together with the parser Świgra used to generate the trees (Section 2) and the valence dictionary Walenty (Section 3). Next, we discuss adapting the parser to Walenty (Section 4) and then adapting the tree- bank (Section 5). 2. The treebank and the parser Składnica is a treebank of Polish (Woliński et al., 2011) built on a sub-corpus sampled from the one million word manu- ally annotated sub-corpus (NKJP1M) of the National Cor- pus of Polish – NKJP (Przepiórkowski et al., 2012). Corpus samples consist of a few sentences each, and they sum up to 20,000 sentences. The primary form of the resource comprises constituency trees generated with the DCG (Pereira and Warren, 1980) parser Świgra (Woliński, 2004; Świdziński and Woliński, 2010) and then manually disambiguated and validated (Woliński, 2010). It was assumed as a construction rule for the treebank that all accepted trees have to be actually gener- ated by the parser. The treebank annotators are not allowed to modify trees in any way nor to provide trees for sentences rejected by the parser. The representation of sentences with- out a proper parse tree has to be corrected by improving the parser’s grammar. This leads to an iterative development of the grammar and the treebank. The grammar feeds the treebank and the treebank documents the …

Content truncated to 500 words