Predicting multi-document comprehension: Cohesion network analysis

Abstract

Theories of discourse comprehension assume that understanding is a process of making connections between new information (e.g., in a text) and prior knowledge, and that the quality of comprehension is a function of the coherence of the mental representation. When readers are exposed to multiple sources of information, they must make connections both within and between the texts. One challenge is how to represent this coherence and in turn how to predict readers’ levels of comprehension. In this study, we represent coherence using Cohesion Network Analysis (CNA) in which we model a global cohesion graph that semantically links reference texts to different student verbal productions. Our aim is to create an automated model of comprehension prediction based on features extracted from the CNA graph. We examine the cohesion links between the four texts read by 146 students and their (a) self-explanations generated on target sentences and (b) responses to open-ended questions. We analyze the degree to which features derived from the cohesive links from the extended CNA graph are predictive of students’ comprehension scores (on a [0 to 12] scale) using either (a) students’ self-explanations, (b) responses to comprehension questions, or (c) both. We compared the use of Linear Regression, Extra Trees Regressor, Support Vector Regression, and Multi-Layer Perceptron. Our best model used Linear Regression, obtaining a 1.29 mean absolute error when predicting comprehension scores using both sources of verbal responses (i.e., self-explanations and question answers).