Is compound chaining the serial-order mechanism of spelling? A simple recurrent network investigation

Abstract

Although considerable progress has been made in determining the cognitive architecture of spelling, less is known about the serial-order mechanism of spelling: the process(es) involved in producing each letter in the proper order. In this study, we investigate compound chaining as a theory of the serial-order mechanism of spelling. Chaining theories posit that the retrieval from memory of each element in a sequence is dependent upon the retrieval of previous elements. We examine this issue by comparing the performance of simple recurrent networks (a class of neural networks that we show can operate by chaining) with that of two individuals with acquired dysgraphia affecting the serial-order mechanism of spelling—the graphemic buffer. We compare their performance in terms of the effects of serial position, the effect of length on overall letter accuracy, and the effect of length on the accuracy of specific positions within the word. We find that the networks produce significantly different patterns of performance from those of the dysgraphics, indicating that compound chaining is not an appropriate theory of the serial-order mechanism of spelling.

Keywords:

• Serial order
• Spelling
• Graphemic buffer
• Simulation
• Compound chaining

Acknowledgments

We greatly appreciate the support provided by NIH (National Institutes of Health) Grant DC 006740 to the second author. In addition, we would like to thank B.W.N. and R.S.B. for their enthusiasm and tireless work on this project and Paul Smolensky, Don Mathis, Manny Vindiola, Michael McCloskey, and the members of the Johns Hopkins University Neural Networks Research Lab for many valuable theoretical and implementational discussions related to this project.

Notes

¹ Chaining theory also predicts that erroneously recalled items should serve as poor cues to the next item in the sequence. To isolate the effect of cue confusability, Henson et al. (1996) controlled for the effect that erroneously retrieved items had on retrieval using a conditional probability analysis. The details of this analysis are beyond the scope of this article.

² Competitive queueing has also been proposed as a theory of verbal short-term memory (Hartley & Houghton, 1996).

³ Note that Jordan–Elman networks (the networks used in this study) do not contain these recurrent connections. In the Jordan–Elman networks, hidden unit activation is a function of the previously produced item (output layer activation) and the state of the hidden units on the previous timestep. It is this latter connectivity that allows the networks to function by compound chaining.

⁴ Given the severity of his deficit, R.S.B. had not been asked to spell many nine-letter words.

⁵ In our modification, the word positions that were combined to fit the 5 standardized positions alternated symmetrically around the middle.

Table

⁶ The V-score effectively measures the length of the shortest arm of the V—the V-score is positive if the short arm reverses the direction of the longer arm and negative or 0 otherwise. Because the depth and asymmetry measures are summed together, it is possible for two curves of very different shapes to receive the same V-score—a deep but asymmetric curve with a short right arm may produce the same V-score as a very shallow but perfectly symmetric curve. Since the difference between these two types of curve may hold theoretical significance, the V-score is not a good all-purpose measure of the bow shape. It does provide a rough measure of the shape, however, and is thus sufficient for our purposes. We gratefully acknowledge an anonymous reviewer for raising this issue.

⁷ These values ensured that the small and large networks carried approximately the same amount of representational load—about 40–50 characters in the training list (including start and stop symbols) per each unit in the hidden layer.

⁸ It is not immediately clear why the error rate for Position 2 in the competitive queueing simulation differs from the other positions. One possibility is that the data for Position 2 are drawn from a relatively small pool of errors. While Positions 3–8 each had a pool of more than 300 word-external substitution errors (with more than 5,000 errors at Position 8), only 26 word-external errors were made at Position 2. The overall pattern is clear, however.

⁹ Because of its architecture, the letter units in the CQ network cannot deviate from their trained values on the first timestep—this is the reason that there is no distortion on the first position. Data from Positions 3–8 were included in the analysis in order to make the two networks more comparable.