In: Kokinov, B., Karmiloff-Smith, A., Nersessian, N. J. (eds.) European Perspectives on Cognitive Science. © New Bulgarian University Press, 2011 ISBN 978-954-535-660-5
Short-term Memory for Emotional Words: Separating the Effects of Emotionality and Semantic Relatedness Winston D. Goh (
[email protected]) Weiying Hu (
[email protected]) Department of Psychology, National University of Singapore, 9 Arts Link, Singapore 117570, Singapore
Abstract Facilitation effects of emotionally salient words in short-term memory have been found in a number of studies. However, there may be a confound of semantic relatedness with emotionality in these previous studies. The present study investigated the extent to which emotionality facilitates immediate memory when emotional and neutral words were presented in the context of homogeneous and heterogeneous semantic categories. Evidence of the emotional facilitation effect was only found in the heterogeneous conditions. Emotional salience did not facilitate memory above and beyond the facilitation provided by semantic categorisation. The results suggest that semantic relatedness may be a more salient aid to encoding and retrieval processes than emotionality.
Introduction Recent research suggests that people remember emotional words better than neutral words in immediate memory tasks (Hadley & MacKay, 2006; Monnier & Syssau, 2008). The emotional facilitation effect found in short-term memory for words is consistent with findings observed using long-term memory paradigms such as free recall (e.g. Ferre, Garcia, Fraga, Sanchex-Casas, & Molero, 2010). However, Talmi and Moscovitch (2004) have cautioned that the extent to which emotionality is the locus of the facilitation effects has to be separated from the effects of semantic relatedness. It is well established that memory for word-lists that are semantically categorised is better than memory for lists that are uncategorised; in both the short-term memory (e.g. Goh & Goh, 2006; Poirier & St. Aubin, 1995) and long-term memory (e.g. Mandler, 1967) literature. The present study is an attempt to tease apart the effects of emotionality and semantic categorisation in short-term memory using an immediate recall task. The words used in studies of emotionality effects on memory are typically manipulated by ensuring that emotional words are equated with neutral words in properties such as word length, frequency, and concreteness, but differ on arousal and valence. Talmi and Moscovitch (2004) agued that this may not be enough as emotional words inherently tend to be more inter-related compared to neutral words. For example, Strauss and Allen’s (2008) database of categorisation ratings for emotional words includes 7 categories of emotions, whereas non-emotional categorisation norms such as Overschlde, Rawson, and Dunlosky (2004) include far more categories. A cross-reference of the words used in Monnier and Syssau
(2008) with association norms (Nelson, McEvoy, & Schreiber, 1998) showed that their pleasant words such as friend and love tend to have one or the other as associates and also shared associates such as happy. Conversely, very few of their neutral words were associated with other neutral words. The finding that pleasant words are better remembered than neutral words therefore cannot be unequivocally attributed to differences in emotionality as they could also arise from differences in semantic relatedness. Hadley and MacKay (2006) attempted to control for semantic relatedness by using foodstuffs and cooking related terms for their neutral words as a basis for comparison to their taboo words, which should form a highly inter-related category. It was not possible to cross-reference these words with the Nelson et al. (1998) norms as taboo words were not included in the latter. Hadley and MacKay acknowledged that their word-lists were not explicitly equated a priori, and instead relied on the null findings between taboo words and neutral words in their first two experiments using pure lists (where the to-be-remembered lists exclusively comprised either taboo or neutral words) to suggest that relatedness did not differ among the words. However, it is possible that a null result could arise from greater relatedness within the neutral words cancelling out the effects of greater emotionality in the taboo words. Using a delayed free recall paradigm, Talmi and Moscovitch (2004) were the first to specifically test the hypothesis that emotionality enhancement may be explained by differences in semantic relatedness among to-be-remembered word-lists. They compared recall for negatively valenced words to two conditions of neutral words, in which the neutral words were either randomly selected from different categories in one condition, or were all from one or two specific categories, such as kitchen, music, and driving, in the other condition. Semantic relatedness was equated for the negative words and the neutral categorised words on the basis that there was no difference in the average subjective ratings of pairwise associations between words obtained from pilot studies. The words from the random neutral condition had lower association scores than the negative and categorised neutral words. Results showed that when compared to the random neutral words, negative words were recalled significantly better, replicating the classic facilitation effect. However, when compared to the categorised neutral words, negative words were not recalled better, and in at least one experiment were in fact reliably worse. These results
strongly suggest that emotionality did not increase performance above and beyond the enhancement arising from semantic relatedness and other organisational processes that may come into play in memory for categorised lists. Although Talmi and Moscovitch’s (2004) findings underscore the importance of controlling for semantic relatedness, their results do not conclusively rule out the operation of emotionality processes in memory, something which the authors themselves acknowledged in their discussion. We would argue that their design was inadequate in determining if there were effects of emotionality above and beyond those of semantic categorisation (or vice-versa) as they did not fully cross these two factors. Although average pairwise associations between words in the emotional and categorised neutral conditions were matched, it was unclear if the number of categories in both conditions were similar. For example, the negative words in the emotional condition could have come from several negative categories such as anger, fear, anxiety, sadness etc. whereas the words from the categorised neutral condition were from only one category. Hence, it is entirely possible that the categorised neutral condition was relatively more homogeneous, in terms of category membership, than the emotional condition. As such, this may have boosted performance in the categorised neutral condition, or cancelled out any emotionality effect observed in the emotional condition. One question that also arises is whether emotional facilitation effects would emerge if the words were in a context of low-relatedness. Another possible reason why there were null effects when relatedness was controlled is that semantic relatedness may be a more salient organisational aid for encoding or retrieval processes than emotionality. Hence, in the context of matching relatedness at a high level (by choosing highly related items for both emotional and neutral words), relatedness may swamp the effects of emotionality. Would emotionality effects be evident at a low level of relatedness? In past research on semantic categorisation effects in short-term memory (Goh & Goh, 2006; Poirier & St. Aubin, 1995), homeogeneous word-lists, where all within-list words were from the same category (e.g. vehicles) were better recalled than heterogeneous word-lists, where each word was from a different category. The present research adopted a similar manipulation of semantic categorisation to tease apart the effects of emotionality from the effects of semantic relatedness. The emotional words (positive, negative, and neutral) were crossed with the nature of the word-lists (homogeneous and heterogeneous) in order to determine any additive as well as interactive (which would not be possible without a fully crossed design) effects of emotionality and semantic relatedness.
Method Participants Twenty-three undergraduates from the National University of Singapore participated in the study for course credit.
Design A 2 (Semantic Category: heterogeneous, homogeneous) x 3 (Emotionality: negative, neutral, positive) fully within-subjects design was used. However, due to constraints in selecting the stimuli from available norms of emotional words, it was necessary to analyse the negative and positive conditions separately in manipulating the word properties as well as reporting the results, as explained below.
Materials The Strauss and Allen (2008) and Overschelde et al. (2004) norms were consulted for selecting the different categories to be used in the homogeneous conditions. For negative words, three different categories were selected: anger, anxiety, and fear. For positive words, a problem for selection was that there was only a single category of happiness in the Strauss and Allen norms. As such, two different sets of neutral words were selected: one with three categories for comparison with the negative words, and one with a single category for comparison with the positive words. The three categories: unit of time, unit of distance, and part of speech, were selected from the Overschelde et al. norms; the single category: countries, was selected from the ISO 3166-1 official standard of countries (International Organisation for Standardisation, 2008). For the heterogeneous conditions, negative, neutral, and positive words were selected from the above norms as well as from the Affective Norms for English Words (ANEW) database (Bradley & Lang, 1999), with the general constraint that they were not from any of the categories selected for the homogeneous conditions and had few or no associations with other target words as determined by the Nelson et al. (1998) association norms. An initial set of 160 words (of which 85 were from the ANEW database) were rated on valence and arousal, using Bradley and Lang’s methodology, by 20 different participants who were from the same population as the experimental sample, but who did not participate in the experiment, in order to collect local emotionality ratings. The correlations between the ANEW and local ratings of the 85 ANEW words were .97 and .62 for valence and arousal, respectively, which provided confidence that the local ratings were generally consistent with the original ANEW ratings. A final group of 126 words comprising seven sets of 18 words each were chosen for the experiment. Four sets were for the homogeneous conditions (negative, 3-category neutral, positive, 1-category neutral) and three sets were for the heterogeneous conditions (negative, neutral, positive). The characteristics of the seven sets are summarised in Table 1. As the positive homogeneous (1 category) and negative homogeneous (3 categories) conditions and the associated neutral conditions could not be made identical, separate two-way analyses of variances (ANOVAs) were performed for the positive and negative conditions and their relevant neutral control groups to ascertain that the words have been properly selected. One was done for the four conditions
Table 1: Word Properties of the Word Sets in the Experimental Conditions Positive Heterogeneous Homogeneous
No. of Syllables Log Frequency Concreteness Association Density Valence Arousal
Negative Heterogeneous Homogeneous
M
SD
M
SD
M
SD
M
SD
M
SD
Neutral Homogeneous (3 categories) M SD
2.00
0.69
1.78
0.48
1.83
0.79
1.72
0.58
2.00
0.84
1.94
1.06
2.17
0.62
9.18
0.91
9.19
1.29
9.35
1.11
8.69
1.27
9.52
1.26
9.02
1.27
8.90
1.07
2.72
1.24
2.94
1.97
3.03
1.28
2.30
1.43
2.87
1.37
3.64
1.31
-
-
0.28
0.58
1.17
1.25
0.44
0.71
1.28
1.64
0.11
0.32
1.56
1.54
-
-
7.53 6.28
0.69 1.02
7.57 6.03
0.37 0.60
2.42 5.52
0.73 0.75
2.66 5.77
0.36 0.51
5.03 4.83
0.39 0.63
5.01 4.62
0.18 0.28
5.41 5.13
0.55 0.48
formed by the 2 (heterogeneous vs homogeneous) x 2 (positive vs. neutral) design; and another for the four conditions formed by the 2 (heterogeneous vs. homogeneous) x 2 (negative vs. neutral) design. Results showed that words were matched on syllable length and log HAL frequency as obtained from the English Lexicon Project (Balota et al., 2007), and concreteness from the Nelson et al. norms, all Fs < 1.5. Only the valence and arousal ratings of the words differed across the emotionality conditions. The valence of negative words (M = 2.54, SD = .58) were lower than the corresponding neutral words (M = 5.02, SD = .30), F(3, 68) = 175.34, MSe = .21, p < .01. The valence of positive words (M = 7.55, SD = .55) were higher than the corresponding neutral words (M = 5.22, SD = .51), F(3, 68) = 123.34, MSe = .27, p < .01. The arousal of negative words (M = 5.64, SD = .65) were higher than the corresponding neutral words (M = 4.72, SD = .49), F(3, 68) = 16.70, MSe = .324, p < .01. The arousal of positive words (M = 6.15, SD = .84) were also higher than the corresponding neutral words (M = 4.98, SD = .57), F(3, 68) = 17.18, MSe = .51, p < .01. This pattern of results indicates that the selected words were appropriately chosen for the manipulation of emotionality. Negative words had low valence and high arousal; positive words had high valence and high arousal; whereas neutral words had mid valence and low arousal. The association density of the words was calculated by obtaining the number of times the target word appeared as an associate of another target word within the word set. The associates of each target word were obtained from the Nelson et al. norms. For the negative conditions, a two-way ANOVA revealed a main effect of semantic categorisation, F(1, 68) = 16.49, MSe = 1.42, p < .01. The homogeneous negative condition (M = 1.42, SD = 1.57) had higher association density than the heterogeneous negative condition (M = .28, SD = .57). For the positive conditions, a one-way ANOVA was run across the homogeneous positive, heterogeneous positive and heterogeneous neutral conditions. This was because the association density of the homogeneous neutral words (countries) was not available in Nelson et al. and thus this condition was not available for analysis. The ANOVA was significant, F(2, 51) = 8.80, MSe = 0.66, p < .01, and paired comparisons
Heterogeneous
Homogeneous (1 category) M SD
showed that the homogeneous positive condition (M = 1.17, SD = 1.25) had significantly higher association density than both the heterogeneous positive condition (M = 0.28, SD = 0.58) and the heterogeneous neutral condition (M = 0.11, SD = 0.32), ps < .001. The heterogeneous positive and neutral conditions were not significantly different from each other, p > .05. This pattern of results shows that the degree of semantic association among words across the three conditions of emotionality was carefully controlled. Any difference between these conditions can then be solely attributed to the manipulation of emotionality. Only the semantic categorisation factor differed in the degree of semantic association, where homogeneous categories had more within-set associations than heterogeneous categories. This was intended as the words in the homogeneous categories were all from the same semantic category.
Procedure At the beginning of each trial, a READY prompt was shown on the computer monitor for two seconds. Six words were then shown one at a time centred on the monitor at a rate of one second per word. Participants were then given 20 seconds to recall the words in the order they were presented, by writing them down in prepared answer booklets. Across participants, the 18 words within each of the seven word-sets (that correspond to the experimental conditions) were randomly divided into three 6-word lists, except for the homogeneous negative and homogeneous neutral 3-category sets, whose 6-word lists were already fixed according to the selected categories (anger, unit of time, and so on). The order of presentation of the 21 6-word lists and the within-list presentation order were randomised across participants.
Results The separate two-way ANOVAs for the negative and positive conditions that were performed in analysing the word properties was also adopted. Serial recall scoring, where responses were scored only if both item and serial position were correct, and item recall scoring, where only item was considered, produced similar patterns of results.
We only report the results of the latter as summarised in Figure 1.
.12) were better recalled than neutral words (M = .62, SD = .12), F(1, 22) = 13.35, MSe = .01, p < .01. This was evidence for a facilitation effect of positive emotionality when all words had been controlled for semantic category. At the homogeneous condition, recall of neutral words (M = .73, SD = .12) did not differ significantly from positive words (M = .71, SD = .11), F < 1. For neutral words, those in the homogeneous condition were better recalled than those in the heterogeneous condition, F(1, 22) = 17.78, MSe = .01, p
