Automatic discourse structure generation using rhetorical structure theory

This tiiesis addresses a difficult problem in text processing: crealing a System lo automalically dérive rhetorical structures o f text. Allhough thè rhelorical structure lias proven to be useful in many fields o f text processing sucli as text summarisation and information extraction, Systems that auiomalically generate rhetorical structures with high accuracy are difficult to find. This is bccause discourse is one of the biggest and yet least well defined arcns in linguistics. A n agreement amongst researchcrs on the best method for nnalysing thc rhetorical structure of text lias not been found. ' . This thcsis focuscs on investigaliug a method lo generate the rhetorical structures of text. By exploiting différent cohesive devices, it proposes a method to recognise rhetorical relations belween spans by checking for thc appearanec o f thèse devices. Thèse factors include eue phrases, noun-phrase eues, verb-phrase eues, référence words, time références, substitution words, ellipses, and syntaclic information. The discourse analyser is divided into tvvo levels: sentence-level and text-level. The former uses syntactic information and eue phrases to segment sentences into elementary discourse units and to generate a rhetorical structure for each sentence. The latter dérives rhetorical relations between large spans and theii replaces each sentence by its corresponding rhetorical structure to produce the rhelorical structure o f text. The rhetorical structure at the text-level is derived by selecting rhetorical relations to connect adjacent and non-overlapping spans to form a discourse structure that covers the entire text. Constraints o f texlual organisation and textual adjacency are effectively used in a beam search to reduce the search space in generating such rhetorical structures. Expèriments carried out in ihis research rcccivcd 89.4% F-score for the discoursc segmentation, 52.4% Fscore for thc senlence-levcl discourse analyser and 38.1% F-score for the final output o f the System. Il shows that this approach provides good performance cumparison with current research in discourse.