Introducing the Prague Discourse Treebank 1.0

International Joint Conference on Natural Language Processing, pages 91–99, Nagoya, Japan, 14-18 October 2013. Introducing the Prague Discourse Treebank 1.0 Lucie Poláková, Jiří Mírovský, Anna Nedoluzhko, Pavlína Jínová, Šárka Zikánová and Eva Hajičová Charles University in Prague Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Czech Republic {polakova|mirovsky|nedoluzko|jinova|zikanova|hajicova}@ufal.mff.cuni.cz Abstract We present the Prague Discourse Treebank 1.0, a collection of Czech texts annotated for various discourse-related phenomena "beyond the sentence boundary". The treebank contains manual annotations of (1), discourse connect- ives, their arguments and senses, (2), textual coreference, and (3), bridging anaphora, all carried out on 50k sentences of the treebank. Contrary to most similar projects, the annota- tion was performed directly on top of syntactic trees (from the previous project of the Prague Dependency Treebank 2.5), benefiting thus from the linguistic information already exist- ing on the same data. In this article, we present our theoretical background, describe the an- notations in detail, and offer evaluation num- bers and corpus statistics. 1 Introduction and Motivation Large collections of gold standard language data are known to build an indispensable base for many NLP algorithms. Reliable morphological tagging and syntactic analysis (phrasal or de- pendency) are nowadays quite a standard inform- ation in language corpora released all over the world. With the gradually increasing interest in modeling discourse structure or using various discourse features1 in different NLP tasks (ana- phora resolution, summarization, MT), also the development of resources aimed at representing various discourse-related aspects has gained on importance. Moreover, both theoretical discourse research and NLP algorithms can benefit from a reliable multi-dimensional analysis of the data (Webber et al., 2003, Stede, 2004). There are already several elaborate theoretical concepts on 1 The term of discourse in this paper is used in two mean- ings. The broader interpretation is roughly equal to text (as in discourse structure, discourse features or discourse co- herence) whereas the narrower sense denotes semantic rela- tions between propositions (as in discourse relations). discourse coherence brought to life in real-data annotation (see Sections 1.1 and 1.2). Still, it is only in recent years that large-scale corpora with manual annotations of sentential and discourse level phenomena have become available. Even fewer such corpora exist that combine more types of manual discourse-level annotations. In this paper, we present a large-scale manual annotation project for Czech in which, apart from the "standard" analysis of a sentence (morpho- logy, synctactic trees), several discourse phe- nomena are marked, all over the same data: pro- nominal, nominal and zero2 coreference, dis- course connectives (henceforth DCs) and the se- mantic relations they express, and the associative relations of the so-called bridging anaphora. The paper is structured as follows: In Sections 1.1 and 1.2, brief overviews of recent projects concerning discourse relations and coreference + bridging anaphora are described, respectively. In Section 2, data and tools used in Prague Dis- course Treebank (PDiT) are introduced. Sec- tion 3 describes the annotation scenario and is followed by evaluation of the project in compar- ison with …

Content truncated to 500 words