Multilingual Extension of PDTB-Style Annotation: The Case of TED Multilingual Discourse Bank

Multilingual Extension of PDTB-Style Annotation: The Case of TED Multilingual Discourse Bank Deniz Zeyrek1, Am´alia Mendes2, Murathan Kurfalı1 1Informatics Institute, Middle East Technical University, Ankara, 2Center of Linguistics, University of Lisbon, Lisbon dezeyrek@metu.edu.tr, amaliamendes@letras.ulisboa.pt, kurfali@metu.edu.tr Abstract We introduce TED-Multilingual Discourse Bank, a corpus of TED talks transcripts in 6 languages (English, German, Polish, European Portuguese, Russian and Turkish), where the ultimate aim is to provide a clearly described level of discourse structure and semantics in multiple languages. The corpus is manually annotated following the goals and principles of PDTB, involving explicit and implicit discourse connectives, entity relations, alternative lexicalizations and no relations. In the corpus, we also aim to capture the character- istics of spoken language that exist in the transcripts and adapt the PDTB scheme according to our aims; for example, we introduce hypophora. We spot other aspects of spoken discourse such as the discourse marker use of connectives to keep them distinct from their discourse connective use. TED-MDB is, to the best of our knowledge, one of the few multilingual discourse treebanks and is hoped to be a source of parallel data for contrastive linguistic analysis as well as language technology applications. We describe the corpus, the annotation procedure and provide preliminary corpus statistics. Keywords: discourse, parallel, multilingual corpus 1. Introduction Over the last decade, there has been a growing interest in parallel corpora for linguistic research or language tech- nology applications, such as Cettolo et al. (2012), Tiede- mann and Nygaard (2004). While most of the existing resources are annotated at the word or syntactic level, as in Bentivogli and Pianta (2005), Haug et al. (2009), cor- pora enriched with discourse-level annotations are scarce but they exist. For example, Stede et al. (2016) introduce a corpus of parallel argumentative texts (German-English) annotated with respect to RST and SDRT; Popescu-Belis et al. (2012) describe pronoun and connective annotation over Europarl for English and French, and Samy and Gonz´alez- Ledesma (2008) report the development of an annotated corpus of UN documents in English, Spanish and Arabic by adopting a pragmatic perspective. Given the scarcity of discourse-annotated parallel corpora, we believe the com- munity would benefit from new resources which involve various languages. Here we describe TED-Multilingual Discourse Bank (TED-MDB), a corpus of TED talks tran- scripts involving European languages as well as a non- European language, Turkish, annotated at the discourse level following the PDTB approach (Prasad et al., 2014). Our effort is based on years of discourse research and the principles that are being established in the science of anno- tation (Hovy and Lavid, 2010; Ide and Pustejovsky, 2017).1 1.1. Discourse Relations and the Role of Discourse Connectives in a Text Discourse is a unit above the sentence level mainly struc- tured in terms of lexical links, anaphoric relations and dis- 1The researchers and annotators involved in the creation of TED-MDB are members of the consortium initiated by the COST project Textlink (http://textlink.ii.metu.edu.tr). For information regarding the corpus, see http://www.textlink.ii.metu.edu.tr/ted- multilingual-discourse-bank. course relations. In this paper, our focus …

Content truncated to 500 words