Abstract

In this thesis, we will discuss the work we have conducted on the relationship between discourse relations in English documents and their associated text-types. Obtaining an understanding of the text-type of a given document is a step towards identifying its larger discourse schema which, in turn, is instrumental in effectively identifying discourse relations. In order to study the relationship between discourse relations and discourse structures, and the text-type of a document, we have created a corpus of documents belonging to seven distinct text-types, from which we extracted discourse relation annotations using already existing parsers. Utilizing the data obtained, we have studied various ways in which discourse relations and text-types are linked in an effort to better understand how discourse schemas can be identified and subsequently utilized in the automatic extraction of discourse relations. Our experiments have shown that the classification of documents within our seven text-types is still better performed with a bag-of-words approach, but the results obtained with the automatically extracted discourse relations suggest that there is in fact a link between text-types and the use of specific discourse relations. We also found that the various text-types are identified with varying accuracy, with text-types such as 'explanation' and 'report' being harder to identify, regardless of the methods used. Finally, our results also show that the cue phrases used to identify explicitly stated discourse relations are amongst the more informative features of our better performing bag-of-words model, and can be utilized to reduce the feature space of this particular model.