Non-semantic contributions to “semantic” redundancy gain

Abstract

Recently, two groups of researchers have reported redundancy gains (enhanced performance with multiple, redundant targets) in tasks requiring semantic categorization. Here we report two experiments aimed at determining whether the gains found by one of these groups resulted from some form of semantic coactivation. We asked undergraduate psychology students to complete choice RT tasks requiring the semantic categorization of visually presented words, and compared performance with redundant targets from the same semantic category to performance with redundant targets from different semantic categories. If the redundancy gains resulted from the combination of information at a semantic level, they should have been greater in the former than the latter situation. However, our results showed no significant differences in redundancy gain (for latency and accuracy) between same-category and different-category conditions, despite gains appearing in both conditions. Thus, we suggest that redundancy gain in the semantic categorization task may result entirely from statistical facilitation or combination of information at non-semantic levels.

Keywords:

• Redundancy gain
• Semantic processing
• Categorization
• Race model
• Coactivation

Notes

¹By analogy, consider a criminal trial in which two pieces of evidence are provided, each of which would unequivocally lead to a conviction (e.g., a video of the crime taking place, and a confession on behalf of the defendant): though the pieces of evidence are not identical, with reference to the binary decision required (e.g., guilty vs. not guilty) the information obtained from them is the same (namely, that the defendant is guilty).

²It is also worth noting that separate activation models with channel interactions can lead to violations of the race model inequality: see, for example, Mordkoff and Yantis (1991, 1993). Without a specific theoretical account of how such interactions would occur in the semantic categorization task, we do not consider these models further here; however, we do return to the idea of channel interaction models in the General Discussion.

³We classified participants as having no clear preference when their handedness scores were between –30 and 30 (with scores of –100 and 100 indicating complete preference for the left or right hand, respectively.

⁴Mean word frequencies for the two categories (occurrences per million words) were 38.75 for the Body words, and 43.21 for the Building words. The 40 randomly selected words from each category used in the two-category condition had slightly lower frequencies, at 31.03 and 33.80 for Body and Building words, respectively. Neither of the differences between categories was significant (both p > .65).

⁵Percent correct ANOVA were undertaken on both raw and arcsin-transformed data. As these showed qualitatively identical patterns of results in all cases, only the raw data ANOVA are reported.

⁶As well as using the Biederman and Checkosky (1970) comparison between single and redundant-target RTs, we also performed analyses using the less conservative comparison suggested by Miller and Lopes (1988). As both comparisons led to the same patterns of results (except in one case, where noted), we report only the former here.

⁷Specifically, as there were no redundant trials with two items from the same category, an expanded version of the ANOVA used in the one-category condition was inappropriate.

⁸When using the method suggested by Miller and Lopes (1988), redundant responses were significantly faster than single-target responses for the Building participants, t(23) = 2.59, p = .017.

⁹We also performed analyses of percent correct data to determine whether the single-attended location account discounted by Shepherdson and Miller (2014) could account for our data. According to this account, redundancy gain can be a result of participants only attending to a single stimulus location on each trial. These analyses showed that the proportions of correct responses in single-target conditions were significantly higher for Experiment 1 than the model could account for. Consequently, we do not discuss this issue further.

¹⁰Note that, in principle, the single-target comparison conditions for redundant–same and redundant–different conditions are different, because they involve the best-of-two-same versus the best-of-two-different single-target conditions. In fact, these conditions were almost identical, such that the minimal difference is only visible in Figure 5(b).

¹¹We thank a reviewer for drawing our attention to this fact.