Emotional valence and arousal are generally considered the two dimensions that define the structure of emotion (Feldman Barrett & Russell, Reference Barrett and Russell1999; Reisenzein, Reference Reisenzein1994; Russell, Reference Russell2003). Valence describes the extent to which an emotion is pleasant or unpleasant (positive, negative), whereas arousal refers to its degree of activation, that is, how exciting, agitating, or otherwise calming an emotion is perceived (Feldman Barrett & Russell, Reference Barrett and Russell1999). Despite the utility of these dimensions in affect research (Russell, Reference Russell2003), several issues remain unresolved. For example, it is not clear whether these dimensions are distinct. Moreover, no consensus has been achieved on the relation between positive and negative affect. For example, it is not known whether these represent distinct aspects or bipolar opposites of the same continuum (Feldman Barrett & Russell, Reference Barrett and Russell1999).

The relationship between emotional valence and arousal can be represented by a U-shaped function, whereby emotionally valenced words (positive and negative) are higher in arousal compared to neutral words (Bradley & Lang, Reference Bradley and Lang1999; Kanske & Kotz, Reference Kanske and Kotz2010; Võ et al., Reference Võ, Conrad, Kuchinke, Urton, Hofmann and Jacobs2009). Although arousal is intrinsically associated with valence (Kissler, Herbert, Peyk, & Junghofer, Reference Kissler, Herbert, Peyk and Junghofer2007; Scott, O'Donnell, Leuthold, & Sereno, Reference Scott, O'Donnell, Leuthold and Sereno2009), several models of emotion assume that valence and arousal are distinct variables (Reisenzein, Reference Reisenzein1994; Russell, Reference Russell2003). Support for this assumption comes from behavioral and neuroimaging research on emotion word processing (Lewis, Critchley, Rotshtein, & Dolan, Reference Lewis, Critchley, Rotshtein and Dolan2007). For example, Lewis et al. (Reference Lewis, Critchley, Rotshtein and Dolan2007) reported a double dissociation in brain activation between emotional valence and arousal during word processing. The amygdala was sensitive to arousal, whereas activation in the orbitofrontal cortex was influenced by emotional valence. Furthermore, different subregions of the orbitofrontal cortex were modulated by increasing positive or negative valence, apparently supporting a characterization of valence in terms of independent axes (positive, negative), rather than a bipolar continuum.

Emotional variables have an effect on single-word recognition during a variety of cognitive tasks including lexical decision (LD), valence judgment, silent reading, emotional Stroop, and self-referential judgment (Estes & Verges, Reference Estes and Verges2008; Kanske & Kotz, Reference Kanske and Kotz2007; Kissler et al., Reference Kissler, Herbert, Peyk and Junghofer2007; Larsen, Mercer, & Balota, Reference Larsen, Mercer and Balota2006; Lewis et al., Reference Lewis, Critchley, Rotshtein and Dolan2007; Scott et al., Reference Scott, O'Donnell, Leuthold and Sereno2009; Võ, Jacobs, & Conrad, Reference Võ, Jacobs and Conrad2006). Studies employing carefully controlled material generally report a different pattern of processing for emotionally valenced words compared to neutral words: valenced words are responded to faster than neutral words. Nevertheless, other studies found differences between positive and negative words depending on the task at hand, that is, slower LD latencies but faster valence judgments for negative words compared to positive words (Estes & Verges, Reference Estes and Verges2008), or a general slow-down effect for negative words when presented in a unique block, but not when intermixed with positive words, suggesting a carry-over effect of negative valence (Algom, Chajut, & Lev, Reference Algom, Chajut and Lev2004).

The manipulation of single emotion words has been employed in studies investigating emotion processing during text comprehension by means of short stories describing events and the subsequent emotional reaction of the protagonist. These texts typically consist of a context sentence (or paragraph) and a following target sentence containing an emotion word, which matches or mismatches the context (consistency paradigm; Gernsbacher, Goldsmith, & Robertson, Reference Gernsbacher, Goldsmith and Robertson1992). Texts containing emotional information are processed in a qualitatively different way than are texts containing chronological information (Ferstl, Rinck, & von Cramon, Reference Ferstl, Rinck and von Cramon2005): processing of emotional aspects of stories specifically activated the ventromedial prefrontal cortex and the amygdala, whereas processing of chronological aspects involved a bilateral frontoparietal network and the left precuneus.

We know that word recognition is influenced by a myriad of lexicosemantic features including word length, frequency, familiarity, age of acquisition (AoA), imageability, and concreteness, all of which predict word naming and LD latencies (Balota, Cortese, Sergent-Marshall, Spieler, & Yap, Reference Balota, Cortese, Sergent-Marshall, Spieler and Yap2004). These variables need to be taken into account and controlled when manipulating affective variables; otherwise, their effects on performance might be confounded with lexicosemantic effects, as was the case in many studies using the emotional Stroop task (for a review, see Larsen et al., Reference Larsen, Mercer and Balota2006). The same is true for studies manipulating lexicosemantic features; for example, abstract words seem to have more affective associations compared with concrete words (Kousta, Vigliocco, Vinson, Andrews, & Del Campo, Reference Kousta, Vigliocco, Vinson, Andrews and Del Campo2011).

Norms for lexicosemantic variables are usually strongly correlated, making it difficult to attribute variability in word recognition to any one dimension alone (Cutler, Reference Cutler1981); for example, word frequency and AoA are correlated (e.g., Juhasz, Reference Juhasz2005), as are concreteness and imageability (e.g., Paivio, Yuille, & Madigan, Reference Paivio, Yuille and Madigan1968). The quadratic (U-shaped) relationship between affective variables suggests that emotion variables may be unitized into a single cluster. It is also possible, however, that each affective variable will be correlated with different lexicosemantic features.

To date, only a few studies have investigated the relationship between affective and nonaffective word properties. Scott et al. (Reference Scott, O'Donnell, Leuthold and Sereno2009) found early interactions between word frequency and valence in event-related potential components, whereby high-frequency valenced words showed larger amplitude in the P1, N1, and early posterior negativity components compared to high-frequency neutral words. The same authors, as well as Kuchinke, Võ, Hofmann, and Jacobs (Reference Kuchinke, Võ, Hofmann and Jacobs2007), reported faster LD latencies for valenced words over neutral ones only for low-frequency words; high-frequency words, instead, showed faster LD latencies for positive stimuli over negative and neutral stimuli. Nakic, Smith, Busis, Vythilingam, and Blair (Reference Nakic, Smith, Busis, Vythilingam and Blair2006) found very low inferior frontal cortex activation for highly frequent negative words, but still significant for highly frequent neutral words, suggesting facilitated processing for the former and thus no need to recruit this region. Overall, these findings suggest that very salient emotional stimuli are prioritized and more easily processed; their salience cannot be attributed to a single variable but to a combination of variables (high frequency, high arousal, and positive valence), where affective dimensions have interactive contributions.

Knowing exactly which variables correlate with arousal and valence would be informative to the question of whether emotion is a distinct cluster whose effects on performance cannot be simply accounted for by the well-known lexicosemantic predictors.

This question has been partially investigated in previous studies. For example, in order to manipulate the affective and lexicosemantic features of single words, experimenters used large word corpora that include information on features such as length and frequency (CELEX; Max Planck Institute for Psycholinguistics, 2001), and subjective ratings of word properties from naive participants on features such as word familiarity, AoA, and imageability. Extant databases in English do not include ratings of emotion variables (Bird, Franklin, & Howard, Reference Bird, Franklin and Howard2001; Medical Research Council [MRC] database, Coltheart, Reference Coltheart1981; The Bristol Norms, Stadthagen-Gonzales & Davis, Reference Stadthagen-Gonzales and Davis2006), and corpora with affective ratings do not include ratings of lexicosemantic features (Affective Norms for English Words [ANEW]; Bradley & Lang, Reference Bradley and Lang1999). Therefore, nothing is known about the correlations between emotion ratings and lexicosemantic variables. One consequence of this gap is that experimenters must retrieve ratings for the same word from different corpora, which often do not overlap in terms of words sampled and scales used. For example, according to Kousta, Vinson, and Vigliocco (Reference Kousta, Vinson and Vigliocco2009), rated AoA and imageability for English words are available for only about a third of the words contained in the ANEW database.

The primary aim of the present study was therefore to generate a corpus of English words suitable for experiments investigating effects of affective as well as lexicosemantic features on single word recognition and on text comprehension. To ensure homogeneous evaluations across words and variables, 300 words were rated in a within-subjects design for emotional valence, arousal, familiarity, AoA, and imageability. The reliability of the results was tested against extant ratings from other corpora. Imageability was chosen, rather than concreteness, as the former feature has repeatedly been shown to better reflect human picturing activity compared to the latter one, in particular with respect to highly imageable abstract words and emotion words (Altarriba & Bauer, Reference Altarriba and Bauer2004; Paivio et al., Reference Paivio, Yuille and Madigan1968).

The corpus includes approximately 100 neutral words to allow comparison between valenced and nonvalenced items. A subset of words comprised adjectives denoting emotions (e.g., happy, sad, indifferent), which are useful for creating texts describing emotional reactions of the protagonists. Possible pairs of antonyms were included for the emotion word subset as well as for other words (i.e., delighted–disappointed, abandon–adopt; see Appendix A), in order to aid the construction of texts for consistency paradigms (Ferstl et al., Reference Ferstl, Rinck and von Cramon2005; Gernsbacher et al., Reference Gernsbacher, Goldsmith and Robertson1992).

The second aim of the study was to explore whether affective variables are correlated with lexicosemantic ones or whether they form a distinct cluster. The latter case would confirm that emotion effects cannot be simply accounted for by other word features, but rather that they contribute to explain additional variance in performance.

If emotional valence and arousal are distinct dimensions of affect, we should expect to observe differential patterns of correlations between lexicosemantic features and each of the affective variables. An additional prediction was that positive and negative words would differ in rated levels of arousal, based on previous studies on words (Võ et al., Reference Võ, Conrad, Kuchinke, Urton, Hofmann and Jacobs2009) and pictures (Lang, Bradley, & Cuthbert, Reference Lang, Bradley and Cuthbert1999). Correlations between word length, familiarity, AoA, frequency, and imageability were also expected, in line with previous studies (Bird et al., Reference Bird, Franklin and Howard2001; Stadthagen-Gonzales & Davis, Reference Stadthagen-Gonzales and Davis2006).

METHODS

Participants

Eighty-two students from the University of Sussex (71 women, 11 men) took part in the experiment and received course credits or £7.50. They were all native speakers of English, between 18 and 42 years (M = 20.5, SD = 3.98).

Material

Word selection

Some words were selected from preexisting emotion word lists: the Berlin Affective Word List (Võ et al., Reference Võ, Jacobs and Conrad2006) and the Compass DeRose Guide to Emotion Words (DeRose, Reference DeRose2005). These items were supplemented with additional words, including emotions expressed as adjectives, ranging from very positive (e.g., delighted) to very negative (e.g., terrified), through neutral (e.g., apathetic) as well as pairs of antonyms. These pairs represent simple suggestions; the appropriate selection of antonyms strongly depends on the experimental design and on the discourse context in which they are going to be embedded. A total of 525 words were initially categorized as positive, negative, or neutral by two native English speakers; 300 words were subsequently selected with approximately one third in each category of positive, negative, and neutral; semantically ambiguous and very infrequent words were eliminated.

The word corpus

The corpus used in the rating study comprised 300 words, approximately one-third positive, one-third negative, and one-third neutral, 90 adjectives, 118 nouns, 24 verbs, 55 words that could be either verbs or nouns in English, and an additional 13 words belonging to more than two categories. Sixty-one items were classified as concrete and 153 as abstract, according to the definition by Paivio et al. (Reference Paivio, Yuille and Madigan1968): “Concreteness was defined in terms of directness of reference to sense experience” (p. 1). Eighty-six items were emotions or adjectives that can be used to express feelings (e.g., “I am happy”; “I feel brave”); we included these categories in the corpus, as this distinction has proven useful in previous research (Altarriba & Bauer, Reference Altarriba and Bauer2004). Frequency of use per million (spoken and written combined) and length in letters, phonemes and syllables were taken from the web-based CELEX database (Max Planck Institute for Psycholinguistics, 2001).

The questionnaire

Online questionnaires were created using the software Macromedia Dreamweaver MX 2004. For each feature, a definition for subjective rating was given (see Appendix B), together with instructions including two examples of words at the extremes of a 7-point Likert scale. Definitions of familiarity, AoA, and imageability were adapted from the Bristol Norms (Stadthagen-Gonzales & Davis, Reference Stadthagen-Gonzales and Davis2006). The extremes were labeled as follows: the scale for emotional valence ranged from −3 (very negative) to +3 (very positive); arousal, familiarity, and imageability were scaled from 1 (not at all) to 7 (very high), and the AoA scale was labeled using the following age ranges: 0–2, 2–4, 4–6, 6–9, 9–12, 12–16, and older than 16. This type of age labels and the use of 7 points represent a standard method for representing AoA (Gilhooly & Loogie, Reference Gilhooly and Loogie1980; Morrison, Chappell, & Ellis, Reference Morrison, Chappell and Ellis1997). At the right end of each scale, the option “unknown word” was given.

Procedure

The participants were given a URL via E-mail to access the questionnaire and enter their details. Then, they read general instructions first, followed by the specific instructions for the first feature to be rated. The words were then presented, each one at the center of the page immediately followed by the 7-point scale. When all 300 words were rated for one feature, instructions for the next feature rating appeared. In this way, no influence was ensured among ratings for different features for the same word. The order of features and the order of words within each rating task were varied for each participant. Because of their similarity, familiarity and AoA were never adjacent, nor were emotional valence and arousal. The rating task was self-paced and could be completed in one or two sessions within a week. Completion took approximately 1.25 hr.

Data analysis

Overall, fewer than 0.19% of responses were rated “unknown.” Only four words in the corpus (1.3%) generated unknown responses. Means and standard deviations of ratings were calculated for each feature for each word. In addition, a categorical variable “valence category” was created based on the following criteria: words rated from +3 to +1 were categorized as positive, from +0.8 to −0.8 as neutral, and from −1 to −3 as negative. A gap between positive and neutral words, as well as between neutral and negative words was left to reduce ambiguity in categorization.

Reliability analyses were carried out, correlating the ratings of appropriate subsets of words from the current corpus with ratings from other corpora.

In line with previous literature, plotting valence and arousal ratings revealed a quadratic relationship between these variables (see Figure 1). A quadratic regression with valence as predictor and arousal as outcome variable was conducted: the effects of all other lexicosemantic features were partialled out by entering them as predictors in a first step; valence and valence squared were then entered in a second step.

Figure 1.

Emotional valence ratings plotted with arousal ratings.

Partial correlations between pairs of lexicosemantic features (familiarity, AoA, imageability, log frequency, length in letters, phonemes, and syllables) were calculated by controlling the effects of the other lexicosemantic variables as well as arousal, valence and valence squared.

Partial correlations between each of the affective variables and each lexicosemantic feature were calculated by controlling the effects of the remaining variables. Because familiarity and imageability were correlated with valence, as well as with arousal, we decided to also calculate quadratic regressions between valence and each of these variables to explore whether this function more accurately reflects their relationships. All lexicosemantic variables along with arousal were entered in a first step, whereas valence and valence squared were entered in a second step.

RESULTS

The full word list of items from the Sussex Affective Word List (SAWL), their associated variables, and the means and standard deviations of all ratings are too extensive to provide here, but they can be viewed in their entirety online at http://journals.cambridge.org/aps.

Descriptive statistics

The descriptive statistics for all features are shown in Table 1 for the three categories defined according to emotional valence. In addition, Table 2 shows the number of words belonging to each grammatical category broken down by valence category.

Table 1.

Descriptive statistics of rated and objective features by valence category: positive (from 3 to 1), negative (from −3 to −1), and neutral (from 0.8 to −0.8) words

Note: Words that fell in between positive and neutral categories (10 words) or negative and neutral categories (10 words) are not included.

Table 2.

Frequencies of positive, neutral, and negative words broken down by grammatical category

Reliability analyses

The ANEW corpus (Bradley & Lang, Reference Bradley and Lang1999) contains valence and arousal ratings for 113 out of the 300 words used in the current study: 47 words were positive, 42 negative, and 21 neutral; 3 words were in between valence categories. Pearson correlations between the ANEW ratings and the present ratings were highly significant (valence: r = .97, p < .001; arousal: r = .73, p < .001).

The MRC Psycholinguistic database (Coltheart, Reference Coltheart1981) contains familiarity and imageability ratings for 181 words and AoA ratings for 72 words out of the 300 items. Correlations with ratings from the MRC database were highly significant (familiarity: r = .78, p < .0001; imageability: r = .92, p < .0001; AoA: r = .93, p < .0001). Finally, in the Bristol Norms (Stadthagen-Gonzales & Davis, Reference Stadthagen-Gonzales and Davis2006) familiarity, imageability, and AoA ratings for 53 words were available. Despite the lower number of shared words, correlations were high (familiarity: r = .91, p < .0001; imageability: r = .96, p < .0001; AoA: r = .95, p < .0001).

Relationships among word features

Relationship between affective features

In the quadratic regression predicting arousal from valence ratings, the first model, which included all lexicosemantic variables, accounted for 7% of the variance (r ² = .07, r = .26), F (7, 299) = 3.08, p = .004, whereas the second model, which included valence and valence squared, accounted for an additional 59% (r ² = .66, r = .81), F (9, 299) = 61.61, p < .001. The regression line including only significant predictors is the following: estimated arousal = 0.16 × familiarity + 0.16 × AoA + 0.09 × imageability − 0.24 × valence + 0.43 × valence².

A t test revealed that rated arousal was significantly higher for negative words than positive words, t (198) = −8.46, p < .0001.

Correlations among lexicosemantic features

All correlations and their significance levels are reported in Table 3. Familiarity showed negative moderate correlations with AoA (r = −.67), positive moderate correlation with imageability (r = .40), an almost moderate positive correlation with log frequency (r = .35), and a low positive correlation with length in phonemes (r = .13). Familiar words are acquired earlier, are more imageable, and are used more frequently than unfamiliar words. AoA was also moderately negatively correlated with imageability (r = −.57), with early acquired words being highly imageable. Log frequency showed a low positive correlation with length in syllables (r = .12). Finally, length in letters and phonemes were highly positively correlated (r = .70), as well as length in phonemes and syllables (r = .53). No other significant correlations were found.

Table 3.

Correlations among lexicosemantic features (r values) and their significance levels

Note: AoA, age of acquisition.

*p = .05. ***p = .001.

Relationships between affective and lexicosemantic features

Familiarity showed a highly significant low positive correlation with valence (r = .29) and a significant low positive correlation with arousal (r = .14): positive words and arousing words are more familiar than negative and nonarousing words, respectively. Imageability showed positive low correlations with arousal (r = .15) and valence (r = .13), whereby arousing words and positive words are more imageable than nonarousing and negative words, respectively. AoA also showed a low negative correlation with arousal (r = −.13), whereby early acquired words are less arousing. See Table 4 for all correlations and their significance levels.

Table 4.

Correlations (r values) between each affective feature and lexicosemantic features and their significance levels

Note: AoA, age of acquisition.

*p = .05. **p = .01. ***p = .001.

In the quadratic regression predicting familiarity from valence ratings, all lexicosemantic variables and arousal predicted 62% of the variance in the first step (r ² = .62, r = .79), F (7, 299) = 67.38, p < .001. Emotional valence only (not valence squared) was an additional significant predictor in the second model, accounting for an additional 4% of the variance (r ² = .66, r = .81), F (9, 299) = 61.20, p < .001. The regression line is the following: estimated familiarity = 6.94 − 0.73 × AoA − 0.22 × imageability + 0.10 × phonemes + 0.41 × frequency_log + 0.13 × arousal + 0.13 × valence.

A linear function is therefore best suited to represent the relationship between familiarity and valence.

A similar result was obtained when computing the quadratic regression with imageability: the first model predicted 47% of the variance (r ² = .47, r = .69), F (7, 299) = 37.10, p < .001, whereas emotional valence predicted an additional 1% (r ² = .48, r = .69), F (9, 299) = 29.96, p < .001. estimated imageability = 11.25 − 0.75 × familiarity − 1.15 × AoA + 0.25 × arousal + 0.11 × valence.

Given linear correlations of imageability with both valence and arousal, we decided to separately examine them for each valence category. Imageability and arousal were correlated within positive: (r = .45, p < .0001) and negative words (r = .60, p < .0001), but not within neutral items (r = −.07, ns).