Mr. Chips: An Ideal-Observer Model of Reading

The purpose of the model is to examine how basic visual, motor and cognitive constraints influence reading behavior in people with normal and low vision. Legge, Klitz & Tjan (1997) show how this model provides novel accounts of characteristic patterns of normal eye movements in reading, including the presence of regressive saccades, word skipping, and a preferred viewing position in words. The paper also describes how the model provides insights into the reading difficulties encountered by people with visual-field loss resulting from eye disease.

We provide here a brief overview of the model followed by three examples of its behavior. Details can be found in Legge, Klitz & Tjan (1997).

Overview This cartoon illustrates the key elements of the model. Mr. Chips uses three sources of information: Visual data are obtained through a simplified "retina" (lower left in cartoon). The retina can have any sequence of high-resolution slots in which letters can be recognized, and low-resolution slots or scotomas where only spacing information is available. We refer to the set of high-resolution slots as Mr. Chips's "visual span." Mr. Chips has lexical information consisting of a dictionary of all allowable words and their probabilities (lower right in cartoon). Any arbitrary lexicon can be used, but the examples below employ a lexicon consisting of the 542 most common words in English. Oculomotor information is available in the form of statistical knowledge of eye-movement accuracy. The Gaussian curve at the middle right of the cartoon indicates that Mr. Chips's eye movements may be "noisy," that is, of somewhat uncertain length. Nevertheless, when Mr. Chips plans a saccade of length L, he knows the probability distribution of the landing sites.

Mr. Chips reads texts that consist of words drawn at random from the dictionary. The model makes no use of syntax or semantics (yet).

Mr. Chips's task is to read through the text in the minimum number of saccades, identifying all the words sequentially without error.

Mathematically, Mr. Chips uses an entropy minimization principle (upper right in cartoon). At any point, he may have partial information about the current word (some of the letters and/or word length information). He "Makes the saccade that minimizes uncertainty (bits of information) about the current word." He can't leave the current word until he has unambiguously identified it (i.e. its uncertainty is zero.)

You can view three examples of short eye-movement (saccade) sequences from the model. These examples illustrate the behavior of Mr. Chips when his "retina" contains blind spots (scotomas). These animations were prepared by Steve Mansfield.

In addition, you can view four new examples. While the examples below are 1-D, these new examples are 2-D. This provides a more realistic representation of an actual human retinal configuration.

The stationary pattern in the upper left is a schematic of the retina. The visual span has a width of 5 letters with a one-letter-wide scotoma at the center. There are 4-letter-wide zones of low-resolution peripheral vision on each side of the visual span. The line of text at the top shows what Mr. Chips is trying to read. As he identifies words, they appear in italics below his moving retina. The highlighted letters represent information about the word(s) he's currently working on. Dots indicate the presence of unrecognized letters. You are seeing repetitions of a sequence of 10 fixations.

Animation by J. Stephen Mansfield

Notice the presence of a regression in this sequence, that is a short leftward saccade. On the sixth fixation, Mr. Chips has partial information about the word; he knows its length and three of the five letters "w.it.". From his lexicon, this could be "write" or "white." To resolve this uncertainty, he has to make a regression to uncover the second letter of the word.

Regressions are often considered a mark of poor reading, but this example shows that an Ideal strategy results in regressions. In fact, an ideal observer with a central scotoma makes substantially more regressions than an ideal observer with intact central vision. Recent research by Mark Bullimore & Ian Bailey (University of California at Berkeley) has shown that human readers with central scotomas also make an unusually large number of regressions.

This example uses a retina in which islands of vision alternate with scotomas. In 2-D, we could call it a Swiss cheese retina, but we refer to this 1-D version as the Alcatraz retina.

Animation by J. Stephen Mansfield

In the repeating sequence, notice that the saccades have a distinctive long-short, long-short pattern. The long saccade shows half the letters of a new part of the text, and then the short saccade brings the other half into view.

This unusual long, short, long, short, pattern of saccades is quite abnormal. Nevertheless, it's optimal for a retina with this pattern of scotomas.

Animation by J. Stephen Mansfield

In the two previous examples, there was no saccade noise. In this third example, there is Gaussian error associated with the landing positions. Although Mr. Chips takes this noise into account in planning saccades, he sometimes runs into severe difficulty. In this example, when he is trying to resolve the ambiguity in "wo.l." with his Alcatraz retina, he needs to identify the 'r' to distinguish "world" from "would". He hunts around rather frenetically for several fixations trying to resolve the ambiguity, but the saccade noise keeps thwarting him. Eventually, he gets the word and moves on.

These three examples show the emergence of regressions, abnormal saccade strategies, and hunting saccades. Analogous behavior is sometimes observed in people with visual-field loss. Is this behavior maladaptive for people with low vision? Not necessarily. The performance of Mr. Chips indicates that such behavior may represent a good, or even ideal eye-movement strategy, linked to the pattern of field loss.

Send questions about Mr. Chips to legge@umn.edu

Page visited 4586 times.
Last updated Monday, August 17, 1998.
©Copyright 1997. Minnesota Laboratory for Low-Vision Research.
Send questions or comments to Click Here