Doctoral Seminar: Elnaz Davoodi

Speaker: Elnaz Davoodi

Supervisor: Dr. L. Kosseim

Supervisory Committee:
Drs. S. Bergler, D. Dysart-Gale, O. Ormandjieva

Title:  Statistical Text Simplification at Discourse-level

Date: Thursday, April 7, 2016

Time: 14:30

Place: EV 3.309

ABSTRACT

In well-written texts, utterances are connected to each other using discourse relations which allow the reader to understand the communicative intention of the writer. Discourse relations (e.g. CAUSE, CONDITION) can be expressed either implicitly or explicitly. Implicit relations are not signalled using lexical cues such as but, since, because, etc. and must be inferred by the readers. On the other hand, explicit relations are signalled using specific terms called “Discourse Markers”. In order to deliver a message through a coherent text with different levels of complexity, different words and syntactic structures can be used, as well as various discourse-level choices. This is done while preserving the text's informational content as much as possible. Discourse relations, the choice of discourse markers and the position of the discourse markers are examples of discourse-level properties which can affect text coherence and text complexity. Textual discourse-level properties have been shown to be correlated to textual dimensions such as their genre. For example, Webber (2009), Bachand (2014), Davoodi (2016) showed a correlation between the textual genre and the choice of discourse relations. In this research, we propose to develop an approach to automatically induce the most appropriate discourse-level changes in order to reduce the complexity of a text. To achieve our goal, we first investigated the correlation between the discourse-level properties of texts and their complexity level (Davoodi (2016)). Then, we will develop a model that recommends appropriate discourse-level choices in order to reduce text complexity. We will use intrinsic and extrinsic evaluation in order to evaluate the performance of our model.