Using LSA Semantic Fields to Predict Eye Movement on Web Pages

Using LSA Semantic Fields to Predict Eye Movement on Web Pages Benjamin Stone (bpstone@psychology.adelaide.edu.au) School of Psychology, University of Adelaide Adelaide, SA 5005 AUSTRALIA Simon Dennis (simon.dennis@adelaide.edu.au) School of Psychology, University of Adelaide Adelaide, SA 5007 Australia et al., 2001), the CWW process takes some aspects of the web page’s display structure into consideration by grouping screen areas into regions. Also, like the Bloodhound Project, the semantic content of the each web page is evaluated statistically against the web user’s target goals. However, instead of using the WUFIS (Web User Flow by Information Scent) algorithm, the CWW uses Latent Semantic Analysis (LSA) to compare semantic content. Furthermore, similar to the Bloodhound Project’s close relative SNIF-ACT, which is a model based on the ACT-R cognitive architecture (Pirolli & Fu, 2003; Pirolli, 2005), the CWW does not limit the content of its statistical semantic analysis to the documents in the website. To this end, both CWW and SNIF-ACT also incorporate a corpus of documents that is considered to represent a user’s knowledge base. Once a web page has been segmented into regions or sections, the model generates a description of each section, and these descriptions are then compared, using LSA, with the users goals and knowledge base. The section with the highest similarity to these user components is then selected for further analysis. Link texts in the selected section are again evaluated against the web user’s goal and knowledge base using LSA. After this evaluation, the model then follows the hyperlink with the highest utility. Abstract This paper outlines a new method for estimating the visual saliency different areas displayed on a web page. Latent Semantic Analysis is used to calculate Semantic Fields values for any (x, y) coordinate point on a web page based on the structure of that web page. These Semantic Field values were then used to predict eye-tracking data that was collected from 49 participants’ goal-orient search tasks on a total of 1842 web pages. Semantic Field values were found to predict the participants’ eye-tracking data. Keywords: LSA; Semantic Fields; LSA-SF; web pages; eye- tracking; visual saliency. Introduction Combining approaches A review of both the Display-based and Semantics-based research into web user’s visual search of web page hyperlinks has indicated that the user’s search processes are influenced by: text semantics, element position, aesthetic qualities of elements, and environmental learning (Brumby & Howes, 2003, 2004; Chi et al. 2003; Faraday, 2000, 2001; Cox & Young, 2004; Kaur & Hornof, 2005; Ling & van Schaik, 2002, 2004; McCarthy, Sasse & Rigelsberger, 2003; Pearson & van Schaik, 2003; Pirolli & Fu, 2003; Rigutti & Gerbino, 2004; Blackmon, Kitajima & Polson, 2005; Grier, 2005; Pirolli, 2005). As is described more fully below, Semantics-based researchers have, to varying degrees, started to incorporate characteristics of the web-page display into their models. Moreover, several researchers have highlighted the importance of this combined approach to modelling users navigation through web sites (Blackmon et al, 2002; Pirolli and Fu, 2003; Chi et al. 2003; and, Kaur & Hornof, 2005). The Cognitive Walkthrough for the Web (CWW) is a theory-based tool designed to assess the usability of websites (Blackmon, Kitajima & Polson, 2005). To this end, CWW simulates web user’s navigation through a website using the CoLiDeS (Comprehension-based Linked model of Deliberate Search) model. Furthermore, CoLiDeS is based on Kintsch’s Construction-Integration theory of comprehension. CWW approaches the problem of modelling web user’s link following behaviour in a somewhat similar fashion to the Bloodhound Project (Chi et al., 2003). Like the Bloodhound Project’s use of page position to inform calculation of probable link choice (Chi Latent Semantic Analysis (LSA) LSA is a statistical method of textual evaluation that allows the researcher to derive meaning from a set of documents (Landuaer & Dumais, 1997; Landauer, MacNamara, Dennis & Kintsch 2007). Linear algebraic methods, such as Singular Value Decomposition, enable the researcher to determine the semantic similarity between words and sets of words contained within a corpus of documents. In a way, the corpus of documents acts as a knowledge base. For example, the Touchstone Applied Science Associates (TASA) document corpus represents literature that students may have been exposed between grade 3 and the first year of college. Moreover, the research described in this paper uses the TASA corpus as a best approximation to the knowledge-base of the first year university students who have participated in this study. LSA - Semantic Fields (LSA-SF) In this paper, an alternative method of modelling human behaviour in web page environments is reported. The LSA-