An evaluation of the interactive-activation model using masked partial-word priming

Abstract

Predictions from Davis and Lupker's (2006) version of the interactive-activation model (McClelland & Rumelhart, 1981) were tested in four masked priming lexical decision experiments. Ambiguous partial-word primes (i.e., ho#se resembles HOUSE and HORSE) preceded word targets with few neighbours (low-N) or many neighbours (high-N) when the word/nonword discrimination was either easy (Experiment 1A) or difficult (Experiment 1B). In a second experiment, unambiguous partial-word primes (i.e., cl#ff resembles only CLIFF) preceded hermit (i.e., words with no neighbours), low-N, or high-N word targets when the word/nonword discrimination was either easy (Experiment 2A) or difficult (Experiment 2B). The model's predictions are supported by the results for the ambiguous primes, but not by the results for the unambiguous primes, particularly when hermit targets are used. A revised definition of the orthographic neighbourhood of a word and/or different assumptions about the impact of frequency on lexical representations would improve the model's ability to account for the data.

Acknowledgements

This research was supported by Natural Sciences and Engineering Research Council of Canada Grant A6333 to the second author and was submitted in partial fulfilment of the requirements for a Master's degree at the University of Western Ontario by the first author.

We would like to thank Ken Forster and Marc Brysbaert for their comments on an earlier version of this paper. Major portions of this research were presented at the 46^th Annual Meeting of the Psychonomic Society, Toronto, Ontario, Canada, November, 2005, the 14^th Meeting of the European Society for Cognitive Psychology, Leiden, the Netherlands, September, 2005 and the 15^th Annual Meeting of the Canadian Society for Brain, Behaviour, and Cognitive Science, Montreal, Quebec, Canada, July, 2005.

Notes

¹At least part of the reason that the model predicts a neighbourhood density constraint is that high-N targets have more shared neighbours with their related ambiguous primes than low-N targets. Across the 60 targets in Experiment 1, the correlation between N and the number of shared neighbours was r = .59, t(58) = 5.58, p < .001.

²A few of the primes formed words if the # is removed (e.g., ho#se). In order to determine whether these primes produced different priming patterns than primes with the # in a different position, the latency analyses were re-run after excluding all word targets that were preceded by a prime that formed a word if the # was removed, on either related or unrelated trials. A total of six low-N words targets and eight high-N word targets were excluded. None of the important findings of Experiment 1A or 1B changed in this analysis.

³As in Experiment 1, a few of the primes formed words if the # is removed (e.g., #loud). In order to determine whether these primes produced different priming patterns than primes with the # in a different position, the latency analyses were re-run after excluding all word targets that were preceded by a prime which formed a word if the # was removed, on either related or unrelated trials. This led to the removal of two hermit, ten low-N, and nine high-N word targets. As in Experiment 1, this did not change any of the important findings. Hermit word targets still produced essentially the same amount of priming as the other words when easy nonwords were used (Experiment 2A) and still received the largest benefit from the prime when more difficult nonwords were used (Experiment 2B).

⁴An additional post-hoc analysis was conducted to examine whether the results in either experiment could have been affected by the position of the letter that was replaced by a # in the prime. In particular, we wished to look at the impact of replacing the first letter on the size of the priming effects. Research suggests that the first letter in a word may be particularly important for word identification (e.g., Bruner & O'Dowd, 1958; Davis, Perea, & Acha, 2007; Perea, 1998). Thus, primes that do not have their first letter (e.g., #ovie-MOVIE) may be less effective than other primes. This possibility is particularly relevant to Experiment 1 because there was a small difference between the low-N target condition and the high-N target condition in the number of first-letter replacement primes (19 of 30 low-N word targets were preceded by related primes which began with a letter and 23 of 30 high-N word targets were preceded by related primes which began with a letter).

To determine if the presence of the first letter of the target word in the prime affects the size of the priming effect, the word targets in both experiments were divided into two groups: those that had a letter in the first position of the prime and those which had a # in the first position of the prime. For Experiments 1A and 1B, 2 (target neighbourhood size: low-N vs. high-N)×2 (relatedness: related vs. unrelated)×2 (prime first position: letter vs. symbol) ANOVAs were run. For Experiments 2A and 2B, 3 (target neighbourhood size: hermit vs. low-N vs. high-N)×2 (relatedness: related vs. unrelated)×2 (prime first position: letter vs. symbol) ANOVAs were run. The most important finding of these analyses was that none of the interactions with prime first position were significant. Priming effects were not, in general, larger for related primes that maintained their first letter. In 5 of the 10 possible comparisons, primes with a # in the first letter position provided more priming than primes having their first letter. In only 3 of the 10 comparisons, did primes with first letters provide more priming than primes with a # in the first letter position. In essence, this analysis suggests that this factor is of limited importance. Certainly, there is essentially no evidence that one should be concerned that the small difference between the low-N condition (19 of 30 related primes with a letter in the first letter position) and the high-N condition (23 of 30 related primes with a letter in the first letter position) might have artefactually affected the results in Experiment 1.

⁵We thank Marc Brysbaert for pointing this out to us.