The utility of modelling word identification from visual input within models of eye movements in reading

Abstract

Decades of empirical work have shown that a range of eye movement phenomena in reading are sensitive to the details of the process of word identification. Despite this, major models of eye movement control in reading do not explicitly model word identification from visual input. This paper presents an argument for developing models of eye movements that do include detailed models of word identification. Specifically, we argue that insights into eye movement behaviour can be gained by understanding which phenomena naturally arise from an account in which the eyes move for efficient word identification, and that one important use of such models is to test which eye movement phenomena can be understood this way. As an extended case study, we present evidence from an extension of a previous model of eye movement control in reading that does explicitly model word identification from visual input, Mr. Chips (Legge, Klitz, & Tjan, 1997), to test two proposals for the effect of using linguistic context on reading efficiency.

Keywords:

• Computational modeling
• Eye movements in reading
• Visual word identification

Acknowledgements

We are grateful to Gordon Legge for sharing the corpus used in the original Mr. Chips experiments. Portions of this work were presented at the 32nd annual conference of the Cognitive Science Society and the 84th annual meeting of the Linguistic Society of America. The research was supported by NIH training grant T32-DC000041 from the Center for Research in Language at UC San Diego to KB and by a research grant from the UC San Diego Academic Senate, NSF grant 0953870, and NIH grant R01-HD065829, all to RL.

Notes

¹O'Regan originally referred to this position as the convenient viewing position.

²There are also other assumptions made about the word identification process that do not relate to visual input, which encode effects of word frequency and predictability.

³Technically, in SWIFT, the preferred saccade length is different for progressive nonrefixations, progressive refixations, regressive refixations, and regressive nonrefixations.

⁴The model of eye movement control in reading other than Mr. Chips that comes closest to this goal is Glenmore (Reilly & Radach, 2006), which incorporates a connectionist model of letter and word activation that bears some similarity to interactive activation models of word identification (McClelland & Rumelhart, 1981; Rumelhart & McClelland, 1982). Crucially for our purposes, however, Glenmore differs from interactive activation models of word recognition in only having a single letter node for each character position (rather than multiple possible letter identities) and a single word node for each word (rather than multiple possible word identities). That is, the model entertains no other candidate letter or word identities other than the correct one, and thus cannot perform word identification.

⁵See Appendix A for formal details of how this algorithm works.

⁶Note, however, that none of the following simulation results we describe from the Mr. Chips model show sensitivity to the particular visual information within words. It seems reasonable given the model's algorithm for saccade target selection to suppose that the model will, e.g., be more likely to skip the endings of words whose beginnings uniquely identify the word, but simulations to verify this have not been performed.

⁷Although we do not believe that word bigram language models are good approximations of how human readers make use of context (and do not encode the type of longer range context effects that reading researchers typically study), we use them here because of their computational simplicity. Such a model gives a lower bound on the amount of benefit that humans might obtain from context, and is sufficient for understanding the qualitative effect of using context in reading.

⁸The models at the extremes of this range—the 100% model without context and the 90% model with context—were used to demonstrate the landing position results in Figure 1.

⁹Although it is orthogonal to our main point here, in which we only use average saccade size and word skipping rate as indices of reading speed, one may ask to what extent the model's saccade size and skip rate resemble human reading behaviour. Unfortunately, both of these measures vary as a function of a number of variables (e.g., text difficulty), so it is difficult to draw a precise comparison. That said, the values produced by the model do appear to be within the usual human range. Rayner (1998) gives the mean saccade size when reading English to be 7–9 characters, a range some of the models we tested fall into (only those with the lower confidence criteria, and more models with context than without). Regarding human readers’ overall word skipping rates in English, a sample of empirical estimates is given by Rayner and McConkie (1976), who report a rate of 51%, Vitu, O'Regan, Inhoff, and Topolski (1995), who report 42%, McDonald and Shillcock (2003), who report 44%, and Greenberg, Inhoff, and Weger (2006), who report 40%, a comparable range to that of our models (39–46%).

¹⁰It should be noted, however, that Mr. Chips’ algorithm is not necessarily the most efficient solution to the problem. For example, as each saccade is planned to maximize the information obtained about the current word, ignoring any information that might be obtained about the next word, this algorithm can be somewhat short-sighted as a way of minimizing the time required to identify the entire text.