Elicitation of knowledge differences in reading comprehension using latent semantic analysis with multiple semantic spaces.

Abstract

Previous research has proposed Latent Semantic Analysis (LSA) as a model and technique of knowledge representation that represents knowledge differences in single semantic spaces (e.g. Grolier's Academic American Encyclopedia, Landauer & Dumais 1997). In this project, LSA knowledge representations were constructed in multiple semantic spaces to represent user knowledge differences for adaptive information retrieval. Semantic spaces with varying degrees of background knowledge were constructed for two versions of a story that participants had read. The two versions induced either complete or incomplete story comprehension. The results indicated that optimal LSA representations depended on the level of story comprehension: LSA representations that were derived from semantic spaces of any size resembled participants' complete story comprehension but matched incomplete story comprehension only if semantic spaces included sufficient information. Larger semantic spaces captured more background knowledge than smaller spaces (Experiment 2). This led to the conclusion that participants with incomplete comprehension relied more on background knowledge to rate word pair relatedness than in the Solved condition where they relied more on story knowledge. Comparing LSA representations in multiple semantic spaces was found to be a viable means for representing knowledge dependent on a reader's background. Implications of these findings for the representation of user knowledge for automated adaptive information retrieval are discussed.