Technical Comments

Response to Comment on “Orthographic Processing in Baboons (Papio papio)”

Science  07 Sep 2012:
Vol. 337, Issue 6099, pp. 1173
DOI: 10.1126/science.1224939

Abstract

Bains pointed out that some of our nonwords were in fact real words and that an algorithm using only information about single letters and their positions achieves the same level of accuracy as baboons in discriminating words from nonwords. We clarify the operational definition of words and nonwords in our study and point out possible limits of the proposed algorithm.

In his Comment, Bains (1) argues that there are two minor methodological problems with the Grainger et al. study (2), the first being that some of the nonwords were actually rare low-frequency words, such as “blub” and “bosc,” and the second that an algorithm based on letters and their positions could learn to discriminate words from nonwords without using higher-level representations such as bigrams.

The first point would appear to reflect a misunderstanding of the goals of Grainger et al.’s study and the experimental procedure used to achieve these goals. Although we used real English words, we could have used an entirely artificial lexicon, as is sometimes done in animal learning studies (3). What makes a word a word in our study is not its presence in a given dictionary but the fact that, in our experiment, words were repeated whereas nonwords were not (each to-be-learned new word was repeated intermixed with different nonwords and already-learned words until accuracy on the new word reached 80%). To further facilitate word-nonword discrimination, words and nonwords differed in bigram frequency (frequency of co-occurrence of letter pairs in the set of four-letter English words that were used). Nonwords were of low bigram frequency (e.g., “blub” and “bosc”), whereas words were of high bigram frequency. Thus, to a monkey in our experiment, “blub” and “bosc” are nonwords—that is, they are not repeated and have low bigram frequencies. Whether they are real English words outside the monkey experiment is irrelevant, at least to a monkey.

In his second point, Bains makes an interesting theoretical observation. Bains demonstrates that a machine learning algorithm based on letters and their absolute positions, which is also referred to as slot coding (4), can discriminate between the words and the nonwords tested in the Grainger et al. study. This finding is fully consistent with our main conclusion that monkeys use an orthographic code (i.e., letters and their positions) to discriminate between words and nonwords. However, Bains does not report the results of two critical tests: (i) whether his algorithm would actually generalize to novel words in the same way that monkeys do (significantly more “word” responses on the first presentation of word stimuli compared with nonword stimuli) and (ii) whether it captures the orthographic similarity effects with nonwords (figures 3 and 4 of Grainger et al.). In the absence of such key tests, the present simulation results are interesting but incomplete.

References

View Abstract

Subjects

Navigate This Article