Fast phonology and the Bimodal Interactive Activation Model

Abstract

Many computational models of visual word recognition and reading postulate a central role for phonology. None, however, has successfully simulated one key phenomenon associated with fast-acting phonological influences during word recognition: masked phonological priming (e.g., bloo primes BLUE better than blai primes BLUE). The tricky issue for computational models is not only to simulate such masked phonological priming effects, but at the same time to correctly read aloud irregular words. This double challenge constitutes a new benchmark phenomenon: the fast-phonology test. It has been previously shown that the dual route model of reading aloud (DRC) does not pass the fast-phonology test, unless it is assumed that lexical decisions are always made on the basis of lexical phonological activation. Here we show that the Bimodal Interactive Activation Model (BIAM), an extension of the interactive activation model, can pass the fast-phonology test, while maintaining the ability to discriminate between words and nonwords on the basis of orthographic activation alone. The BIAM achieves this by virtue of implementing a fast parallel mapping of letters onto input phonemes rather than output phonemes as in DRC. It is argued that the BIAM provides an improved architecture for a general model of visual word recognition and reading.

Keywords:

• Computational modelling
• Dual route model
• Phonological priming
• Phonology
• Pseudohomophones
• Reading

Acknowledgements

The authors thank Conrad Perry for help with the model implementation and also Marc Brysbaert, Kathy Rastle, and Ram Frost for helpful comments on an earlier version of this work. This research was partly supported by grant ANR-06-BLAN-0337.

Notes

¹Tzur and Frost (2007) have pointed out that, applying Bloch's law, it is prime luminance and prime duration that conjointly determine the overall energy level of the prime stimulus, and therefore determine how efficiently it is processed (see also Frost, Ahissar, Gottesman, & Tayeb, 2003; Jacobs, Grainger, & Ferrand, 1995). Nevertheless, prime duration does provide some indication of how fast phonological processing has occurred relative to an estimate of the total time required to recognise a word at a given level of luminance.

²We fully acknowledge that this extension of the BIAM to reading aloud has obvious similarities with CDP+, the only architectural difference being the distinction between input and output phonemes. We examine the motivation behind this critical distinction in the General Discussion.

³These words were taken from Perry et al. (2007), with the exception of “waltzed” and “glimpsed”. These words have more than four graphemes in the coda position.

⁴Following Coltheart et al. (2001), we used printed frequency counts (from the CELEX data base; Baayen, Piepenbrock, & Gulikers, 1995) to calculate resting-level activations for both whole-word orthographic and whole-word phonological nodes. Resting-level activations were calculated following McClelland and Rumelhart's (1981) formula.

⁵We are well aware of the fact that recent experimental results do not support such coding schemes (e.g., Peressotti & Grainger, 1999; Schoonbaert & Grainger, 2004), but the aim here was to keep the input coding as similar as possible to DRC.

⁶It can be remarked that orthographic primes slow down target processing and increasingly so with increasing SOA. This results from prime stimuli activating lexical representations other than the target word, and these inhibit the target word representation. However, what is critical for the present purposes is the difference in response times between the pseudohomophone prime condition and the orthographic control condition at each SOA.