An Assessment of Explicit Inter- and Intra-sentential Discourse Connectives in Turkish Discourse Bank

An Assessment of Explicit Inter- and Intra-sentential Discourse Connectives in Turkish Discourse Bank Deniz Zeyrek, Murathan Kurfalı Informatics Institute, Middle East Technical University, Ankara, dezeyrek@metu.edu.tr, kurfali@metu.edu.tr Abstract The paper offers a quantitative and qualitative analysis of explicit inter- and intra-sentential discourse connectives in Turkish Discourse Bank, or TDB version 1.1, a multi-genre resource of written Turkish manually annotated at the discourse level following the goals and principles of Penn Discourse TreeBank. TDB 1.1 is a 40K-word corpus involving all major discourse relation types (explicit discourse relations at intra- and inter-sentential positions, implicit discourse relations, alternative lexicalizations and entity relations) along with their senses and the text spans they relate. The paper focuses on the addition of a new set of explicit intra-sentential connectives to TDB 1.1, namely converbs (a subset of subordinators), which are suffixal connectives mostly corresponding to subordinating conjunctions in European languages. An evaluation of the converb sense annotations is provided. Then, with corpus statistics, explicit intra- and inter-sentential connectives are compared in terms of their frequency of occurrence and with respect to the senses they convey. The results suggest that the subordinators tend to select certain senses not selected by explicit inter-sentential discourse connectives in the data. Overall, our findings offer a promising direction for future NLP tasks in Turkish. Keywords: Turkish, discourse relations, intra-/inter-sentential connective 1. Introduction Discourse parsing is a challenging task for NLP. It involves various subtasks, such as discourse connective detection, argument detection and sense prediction. Since the release of discourse treebanks, particularly PDTB (Prasad et al., 2014), work on discourse parsing has gained an impetus. But most of this work is on English or European languages; work on other languages is scarce. Turkish Discourse Bank, or TDB, a multi-genre corpus of written Turkish, has been created with this motivation. It has the goal of providing re- searchers with a clearly defined level of discourse structure and semantics and support NLP and LT research in Turk- ish (Zeyrek and Webber, 2008).1 Following the rules and principles of PDTB, TDB (version 1.0) has annotated 8483 relations made salient by explicit discourse connectives, in- volving the discourse connective itself as well as the two text spans it relates (Demirs¸ahin and Zeyrek, 2017). Discourse relations not only hold between adjacent sen- tences (inter-sententially) but can also exist between clauses within a single sentence (intra-sententially). While TDB 1.0 is a richly annotated resource of explicit discourse connectives, it does not annotate all types of intra-sentential discourse connectives and leaves out senses as well as other major discourse relation types that PDTB annotates. Given the need to create a more complete version of TDB and considering budgetary constraints, we have decided to en- hance a modest portion of TDB 1.0 with new annotations. Thus, TDB 1.1, a 40K-word subcorpus has been created as described in Zeyrek and Kurfalı (2017). The subcor- pus includes a collection of 20 text files (each with 2.000 words) distributed according to the genres covered by TDB 1.0 with the following frequencies: fiction (novel …

Content truncated to 500 words