A Computational Mimicry of the Knowledge Augmentation Process in Comprehension Based Learnin

Reading is one of the most dominate means of modern human learning. The process of reading comprehension has long been studied in psychology, yet no algorithm level model for comprehension is available to-date. In this thesis, we explore a plausible computational model behind a difficult form of comprehension- prose comprehension. We define prose as a sophisticated text rich with complex concepts that has specialized meanings and associations that are subtle, and in general difficult to comprehended from the prose alone. Thus a prose text needs the use of external knowledge or references to comprehend. Readers always breakdown the text into concepts and create knowledge associations among them in a graph. Thus the knowledge space of a prose is represented as disconnected subgraphs with multiple components. We postulate a comprehension engine consisting of two major cognitive processes involved; knowledge induction and distillation. The induction process seeks to connect the knowledge space in particular augmenting the associations by incremental reading external reference texts, finding the highest familiarity knowledge associations among the prose concepts. It also uses ontology engine to find lexical knowledge associations among pairs of concepts. The objective is to obtain a knowledge space graph with single giant component to establish a base model for the prose comprehension. We experiment using a version of Steiner Tree called Terminal-to-Terminal Steiner Tree (TTST) that mimics finding the highest familiarity knowledge associations (links) among the prose concepts through no or minimum number of external concepts. The time complexity of the algorithm is O(C+E). The knowledge distillation process is a reduction process on the connected knowledge space graph. It grades the knowledge associations pathways in the connected knowledge space and selects a set of associations that most helps the comprehension for without overwhelming the cognitive capacity. We suggest to use a novel graph grading scheme based on equivalent electrical circuit (EEC) theory for grading the knowledge associations where the highest delivered current flow between the two concepts can mimic the comprehension efficiency of association pathways in the knowledge space graph. To evaluate the proposed comprehension engine, we conduct an experiment with various selected proses. Similar to human readers, we let the comprehension engine mine reference texts from modern knowledge corpuses Wikipedia and Wordnet. For the experiment, we use Gutenberg Project word frequency data to assign familiarity indexes needed in these algorithms. We measure knowledge gain of human readers at pre-comprehension and post-comprehension stages. The results of the experiment verify the potential efficiency of using comprehension engine for improving the quality of human comprehension and saving time.