SzegedKoref: A Hungarian Coreference Corpus

SzegedKoref: A Hungarian Coreference Corpus Veronika Vincze1,2, Kl´ara Heged˝us3, Alex Sliz-Nagy4, Rich´ard Farkas4 1MTA-SZTE Research Group on Artificial Intelligence vinczev@inf.u-szeged.hu 2Department of General Linguistics, University of Szeged 3Department of Psychology, University of Szeged klarahegedus92@gmail.com 4Institute of Informatics, University of Szeged rfarkas@inf.u-szeged.hu a.sliz.nagy@gmail.com Abstract In this paper we introduce SzegedKoref, a Hungarian corpus in which coreference relations are manually annotated. For annotation, we selected some texts of Szeged Treebank, the biggest treebank of Hungarian with manual annotation at several linguistic layers. The corpus contains approximately 55,000 tokens and 4000 sentences. Due to its size, the corpus can be exploited in training and testing machine learning based coreference resolution systems, which we would like to implement in the near future. We present the annotated texts, we describe the annotated categories of anaphoric relations, we report on the annotation process and we offer several examples of each annotated category. Two linguistic phenomena – phonologically empty pronouns and pronouns referring to subordinate clauses – are important characteristics of Hungarian coreference relations. In our paper, we also discuss both of them. Keywords: coreference, corpus, Hungarian 1. Introduction In order to avoid unnecessary repetitions and redundancy, speakers can use a wide variety of expressions when refer- ring to the same entity or event in the world. Languages usually offer several lexical and grammatical tools for this purpose. One of the grammatical tools to express identity is coreference, which is used when two (or more) linguistic units refer to the same entity/individual in the world. Coref- erence relations are most frequently expressed by pronouns, adverbs and nouns (mostly, nouns denoting gender or po- sition such as girl or sergeant). At the lexical level, it is mostly synonyms that can contribute to lexical variability. In this paper we introduce the SzegedKoref corpus, in which coreference relations are manually annotated. For annotation, we selected some texts of Szeged Treebank. It is the biggest treebank of Hungarian that contains manual annotation at several linguistic layers (Csendes et al., 2005). We present the annotated texts, we describe the annotated categories of anaphoric relations – pronominal, nominal, adverbial and verbal coreference and subtypes of nominal coreference (hypernyms, synonyms etc.), and we offer sev- eral examples of each annotated category. We also mark zero anaphors and pronouns coreferential with subordinate clauses since these are two linguistic phenomena of Hun- garian that deserve special attention from the viewpoint of coreference resolution. 2. Related Work There are several coreference corpora available for many languages, for instance, OntoNotes contains coreference annotation for English, Chinese and Arabic (Weischedel et al., 2011; Pradhan et al., 2007). This database formed the training and test sets of the CoNLL-2011 (Pradhan et al., 2011) and CoNLL-2012 (Pradhan et al., 2012) shared tasks, which aimed at automatic coreference resolution. There is coreference annotation in the DIRNDL and AN- COR Centre corpora, containing German and French spo- ken language data (Muzerelle et al., 2014; Bj¨orkelund et al., 2014). As for Japanese, the corpus NAIST Text contains coreference annotation, together with predicate-argument structure (Iida …

Content truncated to 500 words