Coreferential Relations in Basque: The Annotation Process

Abstract

In this paper we present the coreferential tagging of part of the EPEC Corpus of Basque. Although coreference is a pragmatic linguistic phenomenon highly dependent on the situational context, it shows some language-specific patterns that vary according to the features of each language. Due to the fact that Basque is not an Indo-European language, it differs considerably in grammar from the languages spoken in surrounding areas. We will explain these features and the decisions made in each case. After describing the criteria defined for coreferential tagging in Basque, the annotation process will be explained. Our annotation is based on a morphologically and syntactically annotated corpus that provides us with a manageable environment, in which the specific structures that are part of a reference chain can be more easily identified. A part of the corpus was tagged by two annotators who marked up the same text independently, and by another annotator that acted as judge, solving problems in case of disagreement. All this process has been automatized as a result of previous studies carried out in this field. The automatic detection of mentions (Soraluze et al., in: Proceedings of Konvens, 2012) has provided us with a better working environment, and given us the possibility to build a first significant corpus for a later computational treatment of automatic coreferential resolution.

Deep Cross-Lingual Coreference Resolution for Less-Resourced Languages: The Case of Basque

“…For the next experiments two corpora for coreference resolution are used, the EPEC-KORREF corpus (Ceberio et al, 2018) for Basque, the target language, and the OntoNotes English corpus (Hovy et al, 2006).…”

“…For the next experiments two corpora for coreference resolution are used, the EPEC-KORREF corpus (Ceberio et al, 2018) for Basque, the target language, and the OntoNotes English corpus (Hovy et al, 2006 OntoNotes corpus is an English corpus with text from a variety of domains of more than one million words, with annotated mentions and coreferential relations. We used only newswire (nw), and broadcast news (bn) sets, avoiding conversation sets, in order to have texts of the same domain (around 825K words and 100K mentions).…”

EUSKOR: End-to-end coreference resolution system for Basque

“…With reference to the subtask of mention detection, in this section we establish what mentions we regard as potential ones to be included in a coreference chain based on a linguistically motivated mention classification presented in [37].…”

III. Ikergazte. Nazioarteko ikerketa euskaraz. Kongresuko artikulu bilduma. Ingeniaritza eta Arkitektura

“…Aurreprozesaketaren abiapuntua, EPEC-KORREF (Ceberio et al, 2018) corpusa da, 45.000 hitz eta 12.000 aipamen dituena. Corpus honek aipamenak eta korreferentzia erlazioak eskuz anotatuta ditu, korreferentziaebazpenerako baliagarriak izan daitezkeen beste ezaugarri askorekin batera.…”