Developing the Bangla RST Discourse Treebank

Developing the Bangla RST Discourse Treebank Debopam Das and Manfred Stede University of Potsdam Karl-Liebknecht Strasse 24-25, 14476 Potsdam, Germany {debdas, stede}@uni-potsdam.de Abstract We present a corpus development project which builds a corpus in Bangla called the Bangla RST Discourse Treebank. The corpus contains a collection of 266 Bangla text, which are annotated for coherence relations (relations between propositions, such as Cause or Evidence). The texts represent the newspaper genre, which is further divided into eight sub-genres, such as business-related news, editorial columns and sport reports. We use Rhetorical Structure Theory (Mann and Thompson, 1988) as the theoretical framework of the corpus. In particular, we develop our annotation guidelines based on the guidelines used in the Potsdam Commentary Corpus (Stede, 2016). In the initial phase of the corpus development process, we have annotated 16 texts, and also conducted an inter-annotator agreement study, evaluating the reliability of our guidelines and the reproducibility of our annotation. The corpus upon its completion could be used as a valuable resource for conducting (cross-linguistic) discourse studies for Bangla, and also for developing various NLP applications, such as text summarization, machine translation or sentiment analysis. Keywords: corpus, discourse annotation, coherence relations, Rhetorical Structure Theory, Bangla 1. Introduction Coherence in discourse is, to a large extent, achieved through the use of coherence relations. Coherence rela- tions (also known as discourse relations or rhetorical rela- tions) refer to the semantic or pragmatic relations between text segments representing propositions, such as Contrast or Elaboration. For example, consider the following text fragment1: (1) [The U.S. Coast Guard closed six miles of the Houston Ship Channel, where about 150 companies have operations,] [because the thick, black smoke obscured the area.] [wsj-1309]. In Example 1, there are two text segments (marked by square brackets) which are connected to each other by a Reason relation in which the second segment serves as a reason for the first one. We are interested in creating a discourse-annotated corpus in which the texts are annotated with respect to coherence relations. Corpora with relational annotation have become increasingly useful in discourse studies, and they have been developed in various languages. These corpora have also served as valuable resources for developing various computational applications, such as discourse parsing, text summarization, or argumentation mining, to name a few. We have recently begun to develop a corpus of Bangla texts annotated for coherence relations. We use Rhetorical Structure Theory or RST (Mann and Thompson, 1988) as the theoretical framework of the corpus, and we call the corpus the Bangla RST Discourse Treebank or Bangla 1Example source: RST Discourse Treebank (Carlson et al., 2002). The content inside the square brackets following the ex- ample refers to the file number in the corpus from which the text fragment has been taken. RST-DT. In this paper, we present the corpus development project, describing our annotation schemes and annotation procedure. In addition, we also report on an inter-annotator agreement study, by having the initial subset of the corpus annotated by a team of annotators …

Content truncated to 500 words