Modelling Discourse Relations for Arabic

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 736–747, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Modelling Discourse Relations for Arabic Amal Alsaif University of Leeds Leeds, UK LS2 9JT amalalsaif@yahoo.co.uk Katja Markert University of Leeds Leeds, UK LS2 9JT markert@comp.leeds.ac.uk Abstract We present the first algorithms to automat- ically identify explicit discourse connectives and the relations they signal for Arabic text. First we show that, for Arabic news, most adjacent sentences are connected via explicit connectives in contrast to English, making the treatment of explicit discourse connectives for Arabic highly important. We also show that explicit Arabic discourse connectives are far more ambiguous than English ones, making their treatment challenging. In the second part of the paper, we present supervised al- gorithms to address automatic discourse con- nective identification and discourse relation recognition. Our connective identifier based on gold standard syntactic features achieves almost human performance. In addition, an identifier based solely on simple lexical and automatically derived morphological and POS features performs with high reliability, essen- tial for languages that do not have high-quality parsers yet. Our algorithm for recognizing dis- course relations performs significantly better than a baseline based on the connective sur- face string alone and therefore reduces the am- biguity in explicit connective interpretation. 1 Introduction The automatic detection of discourse relations, such as causal, contrast or temporal relations, is useful for many applications such as automatic summarization (Marcu, 2000), question answering (Girju, 2003), sentiment analysis (Somasundaran et al., 2008) and readability assessment (Pitler and Nenkova, 2008). This task has recently seen renewed interest due to the growing availability of large-scale corpora anno- tated for discourse relations, such as the Penn Dis- course Treebank (Prasad et al., 2008a). In the Penn Discourse Treebank (PDTB), lo- cal discourse relations (also called senses) such as CAUSAL or CONTRAST are annotated. They hold between two text segments (so-called arguments) that express abstract entities such as events, facts and propositions. Annotated discourse relations can be signalled explicitly by so-called discourse connec- tives (Marcu, 2000; Webber et al., 1999; Prasad et al., 2008a) or hold implicitly between adjacent sen- tences in the same paragraph, i.e. are not signalled by a specific surface string. In Ex. 1, the connec- tive while indicates an explicit CONTRAST between the attitudes of John and Richard. In Ex. 2, the con- nective while indicates an explicit TEMPORAL rela- tion. In Ex. 3, an implicit CAUSAL relation between the first and second sentence holds. We indicate dis- course connectives and the two arguments they re- late via annotated square brackets. (1) [John liked adventure,]Arg2 [ while]DC[Richard was cautious]Arg2 (2) [The children were crying loudly]Arg1[while]DC,[their mother was cooking]Arg2 (3) [I cannot eat any dessert.]Arg1 [I have eaten far too much already.]Arg2 Although similar corpora for other languages are being developed such as for Hindi (Prasad et al., 2008b), Turkish (Zeyrek and Webber, 2008), Chi- nese (Xue, 2005) and, by ourselves, for Arabic (Al- 736 Saif and Markert, 2010), …

Content truncated to 500 words