Using RSVP and Eye Movement Recording to Determine Usefulness of Information Content Definitions as Predictor for Reading Speed
23.438, Saturday, 17-May, 8:30 am - 12:30 pm, Banyan Breezeway
Yannik T. H. Schelske1, Tandra Ghose2, Thomas M. Breuel1; 1Department of Computer Science, University of Kaiserslautern, Germany, 2Department of Psychology, University of Kaiserslautern, Germany
It is well established that reading speed is affected by the information-content of words (Rayner et. al). There exist many different definitions of information-content. We used the normal and the maximal reading speed for a given sentence as measures to determine the usefulness of different information-content definitions. The normal reading speed of a participant was determined by recording the eye movements during a self-paced reading task (reading speed was actively controlled by the participant), the maximal reading speed of a participant was determined by adjusting the display duration of words in a staircase-based rapid serial visual presentation (RSVP) task (reading speed was passively controlled by the experimenter). The information-content of the words in the sentences were determined in multiple ways, namely based on transitional probability (McDonald & Shillcock, 2003) calculated from the British National Corpus (BNC), based on contextual predictability calculated from a cloze task (Frisson, Rayner & Pickering, 2005) and based on different n-gram probabilities calculated from the Google Books Corpus (GBC). We found that the eye movements in the self-paced reading task, indicative for the normal reading speed of the subject, were best predicted by information-content definitions which were based on contextual probability (i.e. broad semantic relationship between target word and the whole of the sentence), and that in contrast, display durations in the RSVP task, were best predicted by information-content definitions based on n-gram probabilities (i.e. narrow syntactical relationship between target word and a small word neighborhood). Our interpretation of these findings is that for early stages in the visual recognition process of words more simple characteristics (namely n-gram probabilities) are exploited whereas more complex characteristics namely contextual probability are later accessed to facilitate the recognition. We therefore suggest choosing the appropriate information-content definition to help understand the limitations of the reading process at different stages better.