Detecting knowledge gaps between texts over networked representations

Abstract

Gaps between knowledge sources are interesting to various stakeholders, as gaps might indicate potential misconceptions awaiting correction, or complex/novel knowledge that requires careful delivery or study. Motivated by these underlying values and recognizing text to be an important medium for knowledge exchange and education, this study explores gaps of knowledge encoded in textual sources.

Classical cognitive models of text comprehension have established the mapping of human cognitive procedures and computations to the operations over networked knowledge representations, while the latest achievements in network science and optimization theory have equipped us with powerful representation and computational capacity. Hence, this study develops novel methods for gap measurement and prediction from multiple perspectives.

This dissertation consists of three studies probing the knowledge gap detection problem in respect of symbolic structure, content semantics and ways of integrating both evidence from the knowledge networks. Leveraging a networked representation, the first study proposes metrics that capture the dynamics of symbolic structures of knowledge based on network analysis; the second study quantifies the content semantic distances between knowledge elements by framing a cost optimization problem for semantic transportation; finally, the third study develops deep graph neural networks to integrate all the evidence, local and global, for prediction.

The significance of this study is three-fold: first, it enriches the theories of text comprehension by making the first attempt to frame the idea of a knowledge gap and expose structure- and semantic-related patterns; second, it offers methodological contributions by proposing the means of knowledge network construction and the computational procedures of gap detection through the lens of symbolic structure, content semantics, and prediction by integrating both; third, it implements and empirically validates the proposed methods across multiple datasets of educational question answering discourse, obtains interesting findings, and offers a new set of metrics and computational approaches for automatic text analysis and evaluation. The thesis outputs not only enrich findings in the learning analytics field, but also contribute to educational practices involving the analysis and comparison of learners’ knowledge states.