The role of valence and arousal for phonological iconicity in the lexicon of German: a cross-validation study using pseudoword ratings

ABSTRACT

The notion of sound symbolism receives increasing interest in psycholinguistics. Recent research – including empirical effects of affective phonological iconicity on language processing (Adelman et al., 2018; Conrad et al., 2022) – suggested language codes affective meaning at a basic phonological level using specific phonemes as sublexical markers of emotion. Here, in a series of 8 rating-experiments, we investigate the sensitivity of language users to assumed affectively-iconic systematic distribution patterns of phonemes across the German vocabulary:

After computing sublexical-affective-values (SAV) concerning valence and arousal for the entire German phoneme inventory according to occurrences of syllabic onsets, nuclei and codas in a large-scale affective normative lexical database, we constructed pseudoword material differing in SAV to test for subjective affective impressions.

Results support affective iconicity as affective ratings mirrored sound-to-meaning correspondences in the lexical database. Varying SAV of otherwise semantically meaningless pseudowords altered affective impressions: Higher arousal was consistently assigned to pseudowords made of syllabic constituents more often used in high-arousal words – contrasted by less straightforward effects of valence SAV. Further disentangling specific differential effects of the two highly-related affective dimensions valence and arousal, our data clearly suggest arousal, rather than valence, as the relevant dimension driving affective iconicity effects.

KEYWORDS:

• Iconicity
• sound symbolism
• valence
• arousal

According to de Saussure’s (1916/2011) first principle in linguistics, the relationship between sound and meaning is arbitrary, that is, based exclusively on convention (see also Hockett, 1958). On the other hand, the notion of sound symbolism or iconicity that posits that some aspects of language sound may display a direct relation to the meaning of words has a long tradition – see the Cratylos dialog (Plato 4th century BCE/1892) or classical debates from the early era of Enlightenment (Leibniz, 1765/1996; Locke, 1690/1948).

From a theoretical perspective, Jakobson (1960) states that iconicity goes beyond the syntagmatic relationship between linguistic elements and is based on a phenomenological relation between different sensory modalities (see also Dingemanse et al., 2015; Sidhu & Pexman, 2018 for review). Further, Peirce’s semiotics framework (1974; summarised by Liszka, 1996) distinguishes three types of signs: While the symbol, determined by convention, reflects de Saussure’s (1916/2011) notion of arbitrariness, indices and icons stand apart – involving contingency between the sign and the signified (see Dingemanse et al., 2015 for more recent accounts of fundamental mechanisms underlying phenomena of non-arbitrary sound-meaning relations): Whereas indices (e.g. smoke as a sign of fire) involve systematicity that can be understood as the result of an internalisation of statistical co-occurrences of features in the natural environment (see Spence, 2011), iconicity (in the vocal domain often referred to as sound symbolism) is based on perceptual similarities between signifier and signified (e.g. a photograph), in line with accounts based on cross-modal processing or synaesthetic experience as mediated via particular aspects of the neuro-cognitive architecture (e.g. Bankieris & Simner, 2015; Ramachandran & Hubbard, 2001). Importantly, this classification based on Peirce’s (1974) semiotics also allows for an understanding of the evolution of signs from sensory-motor grounding as put forward by theories of embodiment (e.g. Barsalou, 2008; Glenberg, 2010) towards conventionalised use, a process already described by Jespersen’s holistic theory of language genesis (1922) or Darwin’s (1871) speculations in terms of imitation theory.

A growing body of evidence from quantitative linguistics and empirical-experimental studies suggests – on all levels of neurocognitive research (experience, behaviour, brain activity, and computational models) – that the idealised concept of arbitrariness in natural languages should be complemented by non-arbitrary correspondences between sound and meaning. Modern psycholinguistic research has provided solid foundations for such speculations, particularly in sign languages (e.g. Thompson et al., 2020; Vinson et al., 2015), as well as in gesture (e.g. Dick et al., 2014) or prosody (e.g. Perlman et al., 2015). Regarding spoken language, Köhler’s (1929) groundbreaking study showed that subjects commonly tend to associate artificial words (e.g. Maluma or Bouba) containing back vowels (/o/) and sonorant consonants (/m/) with round shapes, but those containing front vowels (/e/) and plosive consonants (/p/) (e.g. Takete or Kiki) with pointed shapes. Similar findings, generally labelled “bouba-kiki effect” have been replicated with impressive agreement among judges, often exceeding 95%, in languages from unrelated language families (see Bremner et al., 2013; Ćwiek et al., 2022; Ramachandran & Hubbard, 2001) and at early stages of language acquisition such as in 2.5-year-old children (Maurer et al., 2006) using a forced-choice procedure and even in 4-month-old infants (Ozturk et al., 2013) in a typical preferential looking paradigm. In another classical study – regarding size rather than shape – Sapir (1929) collected judgements for non-words concerning the objects’ size they presumably referred to and found back versus front vowels to be associated with larger versus smaller objects, a cross-linguistic phenomenon (Shrum et al., 2012; Taylor & Taylor, 1965) also replicated with infants (Peña et al., 2011). Recently growing interest in the topic has extended reports of sound-symbolic effects to further sensory qualities such as taste (Motoki et al., 2020; Pathak et al., 2020), colour (Cuskley et al., 2019), or spatial relations (Johansson & Zlatev, 2013; Vainio, 2021).

Beyond experimental studies on artificial pseudowords, evidence from corpus linguistic studies suggests statistical regularities between sound and meaning to indeed pervade the entire lexicon – an important argument concerning the relevance of sound symbolism for a given language at a more general level. Regarding, for example, the bouba-kiki effect, Sidhu et al. (2021) present evidence for phonemes associated with roundness to appear more frequently in words referring to roundness (and vice versa) in the English lexicon. Using iconicity ratings, again for English words, Winter et al. (2017) demonstrated higher measures of iconicity to appear more often in words referring to sensory domains. Negative correlations between iconicity ratings and age-of-acquisition (Perry et al., 2015) further corroborate a supportive role of iconicity in first language acquisition, as proposed in the sound symbolism bootstrapping hypothesis (Imai & Kita, 2014).

Interestingly, cross-linguistic studies demonstrate also non-speakers of foreign languages are also sensitive to iconic contiguities in unfamiliar languages (Lockwood et al., 2016; Revill et al., 2014), suggesting early communicative advantages for iconicity in everyday language use (Nielsen & Dingemanse, 2021). Complementing this view, Perniss and Vigliocco (2014) emphasise the role that iconicity plays in the development of language through grounding abstract semantic representations in sensorimotor experience (see Harnad, 1990, for the “symbol grounding problem”).

Of note, the phenomenon of inherent connotative meaning is not limited to perceptual dimensions (e.g. shape or size) but also extends to basic affective features: In line with contemporary dimensional theories of emotion (Barrett, 2006; Russell, 1980), Osgood et al. (1957) suggested affect to build the basis of the semantic space. Using the semantic differential, they identified the dimensions of valence, activation (arousal), and potency, mainly affective in nature, to account for the most significant portion of variance covering the connotative meaning of hundreds of pairs of adjectives. Further, affect, and in particular its communicative component, have frequently been discussed as a possible critical starting point for the evolution of language (Darwin, 1871; Jablonka et al., 2012; Panksepp, 2008; Ramachandran & Hubbard, 2001).

Linking the potential role that affect may play in the development of language – both phylogenetically or individually – or for the structuring of semantics, to the development and structure of the lexicon, iconic sound-meaning relations concerning affect have been shown to be observable in the organisation of the lexicon and, further, to influence the processing of linguistic signs, and its neurocognitive correlates – an effect sometimes referred to as affective phonological iconicity (Gafni & Tsur, 2019; Schmidtke, Conrad et al., 2014).

Accordingly, Heise (1966), using a corpus-based approach to iconicity (see also Sidhu et al., 2021; Winter et al., 2017), found systematic deviations in occurrences of phonemes depending on valence and arousal of the words in which these phonemes appear.

As to the artistic use of language – often aiming to elicit emotional experience in the recipient – cognitive poetics used phonemic contrasts to distinguish among groups of texts differing in affective characteristics (Auracher et al., 2010; Fónagy, 1999; Miall, 2001; Whissell, 2000).

From a more general perspective, a number of studies have attempted to establish directly the potential affective impact of specific language sounds: For instance, Ando et al. (2021) showed that individual vowels in Japanese are rated differently on various semantic differential scales, contributing to specific subjective evaluations. Kambara and Umemura (2021) contrasted initial consonants of sound-symbolic words in Japanese and found significant differences in ratings of imageability as well as affective dimensions of valence and arousal depending on the consonants’ voicing. However, the Japanese language comprises the word class of ideophones, characterised by their vivid representation of ideas, actions, or states, through phonetically expressive sound (see e.g. Dingemanse, 2012), a phenomenon observable in many of the world’s languages, but rather underdeveloped in Indo-European languages. Here, reliable effects have so far been reported, for example, for the vowel /i:/ (as in English sea), relating perceived positive valence to overlapping activity of facial muscle activity during the articulation of the sound /i:/ as well as the emotional facial expression of a smile as demonstrated for German (Körner & Rummer, 2022; Rummer et al., 2014; Rummer & Schweppe, 2019), Hungarian (Benczes & Kovács, 2022), English (Yu et al., 2021), but also Japanese (Körner & Rummer, 2023). Nonetheless, these effects are restricted to a fairly small subset of vowels and hardly generalise to the broader scope of the phoneme inventory.

Expanding the focus on a more extensive phoneme inventory, Adelman et al. (2018) demonstrated in a comprehensive corpus-linguistic analysis across five Indo-European languages, that initial phonemes predict words’ valence better as compared to subsequent phonemes. As they found the speed of pronunciation to increase with negative valence and possibly high arousal, they speculate evolutionary adaptation to underly this phenomenon, acting as an early alert system consistent with the automatic vigilance hypothesis (see Fox et al., 2001). To further delve into the potential contiguity of phonological elements and affective perception, Aryani et al. (2018) used a computational approach showing that words’ phonemic structure can predict some degree in the variance of ratings on their semantic affective meaning concerning both valence and arousal. Further, Aryani et al. (2020) demonstrated arousal – defined in terms of acoustic features – to also predict sound-shape mappings as reported in 8 well received previous studies (e.g. Köhler, 1929; Ramachandran & Hubbard, 2001).

Unlike the theoretical position taken by Aryani et al. (2018, 2020), who considered the affective potential of phonemes to influence affective evaluations of words containing these phonemes without being internally related to words’ affective semantic content itself, and, in consequence, used explicit ratings of words’ or pseudowords’ affective sound rather than those of words’ semantic affective content to further explore the role of acoustic features for affective impressions evoked by pseudowords (Aryani et al., 2018), we took a different approach in our “sublexical affective values” line of research (Aryani et al., 2016; Conrad et al., 2022; Schmidtke & Conrad, 2018; Ullrich et al., 2016, 2017), investigating phonological affective iconicity that we see as an inherent feature of the organisation of the vocabulary – marking semantic affective content of words at a sublexical level.

Based on the assumption that affective iconicity permeates the entire lexicon, we used valence and arousal values of all words contained in our normative database for 5695 German words (Schmidtke & Conrad, 2018, extending the BAWL corpus of Võ et al., 2009) to calculate SAV as follows:

Word entries were segmented into sub-syllabic units – rather than single phonemes – based on evidence from both experimental (Brand et al., 2007; Nuerk et al., 2000) and computational simulation research (Jacobs et al., 1998) on language processing underscoring the role of these segments as perceptual units encoding phonology. These sub-syllabic units encompass a nucleus, that is, a mandatory syllabic core characterised by a maximum level of sonority – referring exclusively to vowels in German as in most other, but not all languages – and are optionally accompanied by an onset and/or coda, that is, additional consonants or consonant clusters preceding or following the syllable nucleus, respectively. Affective values of valence as well as arousal of words were then averaged for each resulting segment across all words that these segments occur in to calculate the SAV of segments.

Note that the German language can be considered of special interest for the investigation of affective phonological iconicity for two reasons: First, being a highly transparent orthography with consistent grapheme-phoneme conversion rules, it seems possible to investigate effects presumably arising through phonological processing in rather straightforward ways also using visual stimulus presentation. Second, unlike comparably transparent Indo-European orthographies, as e.g. Spanish, the German language licenses for rather complex consonant clusters combining often up to three different consonants in one syllabic onset or coda. This makes it a priori more plausible that specific consonant clusters could serve as sublexical markers of meaning – information-economy being one reason for the arbitrariness principle in language. Furthermore, many of these complex consonant clusters involve combinations of plosives and sibilants (see words like KATZE (cat), SPRITZE (syringe) or KLATSCH (gossip)) – both of which have been shown to be crucial for affective phonological iconicity (Aryani et al., 2018), and respective effects may sum up when being combined.

In studies of sentiment analysis, Aryani et al. (2016) and Ullrich et al. (2017) demonstrated SAV to serve as reliable predictors of affective impressions evoked by poems (submitted to phonological analysis through the EMOPHON, Aryani et al., 2013). Further, SAV proved to affect online language processing in a number of automatic task paradigms: Schmidtke and Conrad (2018) observed shorter reaction times in a letter search task to graphemes associated with high arousal (occurring more often in words of high arousal than in those of low arousal), alongside notable interactions with the emotional content of the words they were embedded in. Echoing Adelman et al. (2018), the authors proposed the observed enhanced behavioural responses to salient phonological units of negative valence and high arousal to align with fundamental principles of behavioural adaptation, priorizing avoidance of negative consequences, detrimental to the organism, as compared to the potential gains from positive outcomes. Importantly, with regard to the relation between sublexical affective value and formal salience, note that respective SAV correlate with increasing complexity of segments, and, in negative ways, with their frequency of occurrence, that is, the more words a specific segment occurs in, the less likely it is that its SAV deviates from neutral; whereas more “affectively” deviant segments tend to be of higher perceptive salience already at a formal level.

Conrad et al. (2022) observed a reduction in the N400 component of the event-related potential (ERP) for words with a consistent matching between (negative and high arousal vs. neutral) lexical and sublexical affective values during a visual lexical decision task, indicating a processing advantage due to affective iconicity. Ullrich et al. (2016) showed that ERP effects for SAV during visual word recognition are particularly pronounced for those sublexical segments where strong deviations from the neutral range of SAV go along with high perceptual salience, that is for sublexical segments of more complex structure and relatively low frequency of occurrence (ERP effects were attenuated when controlling for salience of sublexical units).

While these empirical findings of SAV effects on language processing strongly suggest that – with regard to valence and arousal – the distribution of phonological units across the German vocabulary is NOT completely arbitrary, further research is needed to investigate the nature of this alleged affective phonological iconicity.

In particular, the questions of whether apparent effects of affective iconicity on language processing, are, on their own, affective in nature, or rather related to attentional processes and perceptive salience, have not been explicitly or sufficiently addressed. To start with, the most straightforward way to test for affective iconicity – concerning both the phonological structure of the lexicon and the processing of language – seems to consist of using the phoneme clusters apparently involved in sound-to-meaning correspondences in the lexicon of German as experimental stimulus material to test for affective impressions – note that above mentioned studies have either only addressed this question via regression models (Aryani et al., 2018), or have only tested experimentally for SAV effects on superficially cognitive task aspects, but not on explicitly affective processing. To overcome this and other shortcomings, the present study investigates affective iconicity in German via a straightforward cross-validation-approach manipulating SAV of German pseudowords used for affective rating studies:

• - Whether potential sound-meaning correspondences with regard to affect emerging across the lexicon can modulate the affective impressions of language users when presented with specific phoneme clusters – investigating the relevance of systematic sound-to-meaning correspondence patterns beyond the organisation of the vocabulary (where SAV was derived from), but at the level of individual affective experience (Experiments 1a/b).

• - To what extent affective phonological iconicity is related to the formal salience of phonological units: Would respective effects be limited to segments with increased perceptual salience – due to increased syllabic complexity and lower frequency of occurrence of SAV deviant from neutral (see Schmidtke & Conrad, 2018; Ullrich et al., 2016) or rather generalise across the whole spectrum of phonological segments (Experiments 2a/b)?

• - The nature of affective mechanisms driving potential iconicity effects, contrasting potentially differential roles for the dimensions valence and arousal (Experiments 3a/b and 4a/b).

To address these questions, we designed a series of pseudoword rating studies with increasing degrees of experimental control (see Table 1 for a synoptic view). We generally predict affective ratings to reflect the probabilities of given phonological segments to occur in words of specific affective content in our database (Schmidtke & Conrad, 2018).

Table 1. Synoptic view of all conducted experiments.

Experimental sets	Independent variable
	SAV Valence	SAV Arousal
Basic manipulation	1a	1b
Controlled for Type frequency	2a	2b
Controlled for alternative affective dimension	3a^a	3b^b
Controlled and rated for alternative affective dimension	4a^b	4b^a

Note. ^abData collected within the same experimental sessions.

Methods

Experiments 1a and 1b

In Experiments 1a and b, we tested for the general validity of our hypothesis, predicting SAV to influence subjects’ evaluations of semantically meaningless pseudoword material. Specifically, if sound-to-meaning correspondences in the database reflected affective phonological iconicity, we expected subjects’ ratings of valence for pseudowords to vary with pseudowords’ mean SAV for valence as well as subjects’ ratings of arousal to vary with pseudowords’ mean SAV for arousal. Note that manipulations involved maximised respective contrasts between conditions of either valence or arousal in order to establish if any effects would occur at all for any of the two dimensions – without controlling for the respective alternative dimension, both being classically highly correlated (see Schmidkte et al., 2014), and the same lack of control is present for formal salience in terms of frequency of occurrence of segments.

Participants

A total number of 24 (15 female; M age = 28, SD age = 7) German native speakers took part in Experiment 1a. For Experiment 1b, 23 participants were recruited (14 female; M age = 28, SD age = 7). Sample sizes were based on comparable experimental approaches using affective ratings for words or pseudowords (18-30 ratings per item; e.g. Aryani et al., 2018; Bradley & Lang, 1994)

All participants of experiments presented in this study were native speakers of German, students from the Freie Universität Berlin, who were remunerated with course credit. All experiments took place in quiet rooms, lasting between 10 and 20 minutes.

Material

All following calculations were based on our database of 5695 German words (Schmidtke & Conrad, 2018), extending the first normative database to provide ratings on the affective dimensions of valence and arousal for German words (Võ et al., 2009), a resource that has been proven reliable through numerous behavioural and neuroscientific studies (see Jacobs et al., 2015 for review). Word entries in the database were rated by a minimum of 30 participants (balancing male and female) on the dimensions of valence, ranging from – 3 (negative) through 0 (neutral) to +3 (positive), and arousal, ranging from 1 (low) to 5 (high), referring to established dimensional models of affect (Barrett, 2006; Russell, 1980). Following Võ et al. (2009), as the German word for arousal (“Erregung”) is sexually connotated, Self-Assessment Manikin (SAM) was used to collect ratings of arousal, that is a nonverbal pictorial assessment technique based on the semantic differential scale developed by Mehrabian and Russell (1974) and adapted by Lang (1980). For arousal, the SAM is composed of a series of 5 cartoon-like figures, ranging from an excited, wide-eyed figure to a relaxed, sleepy figure.

Based on the collected word ratings, we calculated SAV for sub-syllabic phoneme clusters. All words contained in the database were transcribed phonologically using data from the CELEX (Baayen et al., 1996) and segmented into sub-syllabic units of onset, nucleus and coda. SAV was calculated by averaging the affective values of valence as well as arousal for each segment across all words the respective segments occur in. Segments were then categorised according to their resulting SAV – splitting the spectrum into thirds – thus obtaining two (valence/arousal) times three groups of either positive, neutral or negative valence (Experiment 1a), or low, medium or high arousal (Experiment 1b), respectively. For all experiments presented in this study, in order to assure that our pseudoword stimulus material – apart from not possessing semantic meaning – would be perceived as otherwise typical for the German language and free from any formally “strange” features that might trigger affective processes by means of unusualness, and, in order to avoid that our sublexical segments would strongly evoke associations to specific words (see Sulpizio et al., 2021), which could make ratings semantic-driven rather than following phonological iconicity, we only used sub-syllabic phoneme clusters that occurred rather frequently across the 5695 words in our normative database (Schmidtke & Conrad, 2018). To further make sure that potential effects could not be provoked by or attributable to a few single segments with specific presumably idiosyncratic characteristics, we aimed at using the most extensive variety of possible single segments of the onset, nucleus, or coda type within each experimental category. Doing so, our goal is to establish whether affective phonological iconicity would hold across the whole range of the phonological inventory of German, or could – instead, just be considered relevant for a few very specific sounds. After choosing a set of numerically equal (or at least comparable) onsets, nuclei and coda segments for each condition, sub-syllabic clusters were then recombined by permutation of onsets, nuclei and coda within each of these six SAV categories (the same procedure was applied for all experiments contained in this study). From the resulting sets, two separate final sets, one for valence and one for arousal with each a total number of 180 pseudowords were selected – containing 60 items for each condition of positive, neutral and negative valence (Experiment 1a), and again 60 items for each low, medium and high arousal (Experiment 1b). Six different onsets, 4 different nuclei, and 4 different codas per condition were used in the valence list, (minimum type frequencies (occurrences of segments in N words in the database): 474, 612, and 714 respectively, resulting in combinations such as e.g. positive – kient /kiːnt/, neutral – teift /taɪ̯ft/, or negative – kräb /kʁε:b/). Respective numbers for the arousal list were: Five different onsets, 4 different nuclei, and a minimum of 4 different codas per condition (minimum type frequencies: 543, 612, and 714 respectively, resulting in combinations such as e.g. low arousal – mell /mεl/, medium arousal – pien /piːn/, high arousal – speuz /spɔʏz/). For detailed stimulus characteristics, see Table 2.

Table 2. Stimulus characteristics for pseudowords for manipulation of valence (Exp. 1a) and arousal (Exp. 1b).

	SAV’s manipulated for
	Valence (exp. 1a)			Arousal (exp. 1b)
	Negative	Neutral	Positive	Low	Medium	High
SAV Valence	−0.37	−0.04	0.14	0.01	−0.09	−0.29
M (SD)	(0.07)	(0.01)	(0.06)	(0.07)	(0.07)	(0.13)
p	< .001			< .001
SAV Arsoual	3.04	2.91	2.89	2.82	2.92	3.06
M (SD)	0.05	(0.04)	(0.05)	(0.02)	(0.01)	(0.04)
p	< .001			< .001
Type frequency	1043	6526^a	4049	6915	8389	944
M (SD)	(232)	(1798)	(1368)	(499)	(483)	(197)
p	< .001			< .001

Note. M = mean, SD = standard deviation.

^a All post-hoc comparisons result in p < .001.

Procedure

After providing informed consent, participants started with Experiment 1a, where they were asked to rate their subjective impression of the valence of each pseudoword. Stimuli were presented visually and remained on the screen until participants responded. Valence ratings were obtained using a 7-point scale (with verbal anchors ranging from – 3 = very negative over 0 = neutral to +3 = very positive), in accordance with the procedure applied by Võ et al. (2009) or Schmidtke and Conrad (2018). In addition, participants in all experiments were instructed to label a presented pseudoword as “like a word” when it reminded them strongly of one concrete existing word, and no ratings were collected for respective items. Stimuli rated by more than 1 participant as “like a word” were excluded from all analyses to avoid potential semantic effects to influence our data. After evaluating five training trials, the complete set of 180 items was presented in random order, with three self-paced breaks in between. The same presentation mode was applied in all experiments of this study. For Experiment 1b, arousal ratings were obtained using a 5-point scale (with verbal anchors ranging from 1 = very calming to 5 = very exciting), again assisted by the presentation of SAM’s (Lang, 1980) as applied by Võ et al. (2009) or Schmidtke and Conrad (2018).

Statistical analysis

As dependent measures were attained via ratings and the data include nonindependence due to both repeated measures within participants and stimuli (across participants), we applied cumulative link mixed-effects models (ordered logit models) using R (version 4.2.2; R Core Team, 2022) and the “ordinal” package (version 23.12-4; Christensen, 2023). A maximal random effects structure was used, following Barr (2013). The overall model significance is reported based on model comparison to a null-model excluding the fixed effect and containing the same full random effects structure as applied in the full model, using χ²-test statistics (Winter, 2013). Marginal R² values were calculated to quantify the explanatory power of our models based on the fixed effects included. Post-hoc tests were conducted using the “emmeans” package (version 1.8.8; Lenth, 2023) and results are reported on the logit scale (not the response scale), due to the ordinal nature of the response data. Effect sizes (Cliff’s delta) were calculated using the “effsize” package (version, 0.8.1; Torchiano, 2020)

Results

Experiment 1a): Valence

Data and analysis codes for all experiments are available at https://osf.io/y53hn.

The fixed factor of valence (with the levels positive, neutral and negative) as well as random intercepts and slopes for participants and random intercepts for stimuli were entered into a cumulative link mixed-effects model. A total of 3.8% of all items were rated as word-like. Excluding all items receiving 2 or more word-like ratings additionally yielded a total amount of excluded items of 22.8%. Our model yielded an effect for the fixed factor valence, χ²(2) = 29.11, p < .001, R² = .077, with more positive ratings for items of neutral SAV (M = 0.617, SE = 0.151) as compared to negative SAV, (M = −0.813, SE = 0.177), z = −6.60, p < .001, Cliff’s delta = -.38, 95% CI = [-.42, – .33], as well as more positive ratings for positive SAV (M = −0.021, SE = 0.144) as compared to negative SAV, z = −4.03, p < .001, Cliff’s delta = -.22, 95% CI = [-.27, – .18]. Items of neutral SAV also received more positive ratings as compared to positive SAV, z = 4.05, p < .001, Cliff’s delta = .17, 95% CI = [.12, .22] (see Figure 1).

Figure 1. Response patterns for valence ratings for Experiment 1a across different conditions. (a) Estimated response probabilities for each level of valence rating by valence SAV condition (negative, neutral, positive) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each valence condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Experiment 1b): Arousal

The fixed factor of arousal (with the levels low, medium and high) as well as random intercepts and slopes for participants and random intercepts for stimuli were entered into a cumulative link mixed-effects model. A total of 1.7% of the items were labelled as word-like by participants. Excluding all items receiving 2 or more word-like ratings yielded a total amount of 7.5% to be excluded from further analysis. Our model yielded an effect for the fixed factor arousal, χ²(2) = 29.83, p < .001, R² = .108, with higher arousal ratings for items in the high arousal SAV condition (M = 0.586, SE = 0.282) as compared to the low arousal SAV condition, (M = −1.231, SE = 0.282), z = −6.06, p < .001, Cliff’s delta = -.43, 95% CI = [-.47, – .39], as well as higher ratings for items in the medium arousal SAV condition (M = 0.096, SE = 0.273) as compared to the low arousal SAV condition, z = −6.10, p < .001, Cliff’s delta = -.31, 95% CI = [-.35, – .27]. Ratings between the levels of medium and high arousal SAV did not differ, z = −1.70, p = .206 (see Figure 2).

Figure 2. Response patterns for arousal ratings for Experiment 1b across different conditions. (a) Estimated response probabilities for each level of arousal rating by arousal SAV condition (low, medium, high) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each arousal condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Discussion

As predicted, we found negative pseudowords in the valence condition to be evaluated significantly more negative as compared to neutral or positive ones. Pseudoword ratings in the arousal condition similarly confirm our hypothesis, showing low arousing items to be rated significantly lower on arousal as compared to items in the medium or high arousal condition, as well as significantly higher ratings for items high on arousal as compared to the medium condition. These findings correspond well with our initial assumption that phonological characteristics are associated with the perceived valence or arousal in the absence of semantically meaningful referents.

However, in the case of valence, there is a striking misalignment between our predictions concerning valence sublexical affective values and subjects’ ratings for neutral versus positive items, as items in the neutral condition were found to be rated significantly more positive as compared to items in the positive condition.

As this deviation is difficult to account for based on SAV alone, further variables known to influence the affective evaluation of stimuli need to be considered. Regarding valence, the processing fluency hypothesis (Belke et al., 2010; Reber et al., 2004; Snefjella & Kuperman, 2016) states that evaluation processes are based on cognitive fluency, as ease of mental operations may serve as a hedonic marker. Thus, the processing fluency hypothesis would predict more positive evaluations of stimuli with a higher frequency of occurrence, being more familiar and easier to process. Analysis of our stimulus material indeed revealed that segments of the neutral condition occurred in more words in the database as compared to positive or negative items (see Table 2), and had thus to be considered sublexical segments of relatively higher frequency or familiarity, which might have caused the otherwise counterintuitive finding of more positive valence ratings for neutral SAV.

Experiments 2a and 2b

Experiments 2a and 2b were designed to test again for SAV effects of valence and arousal, after controlling for the confound between SAV and formal sublexical salience, respectively frequency or familiarity present in the material of Experiments 1a and b, where we had aimed at providing an initially most powerful contrast between SAV conditions to set the seminal findings for our study.

Participants

Based on the result of Experiment 1a, a simulation-based power analysis was performed using the R-package mixed power (Kumle et al., 2021), based on linear mixed-effects models using the lmerTest package (version 3.1-3; Kuznetsova et al., 2017), as packages for power analysis of cumulative link mixed-effects models are not readily available yet. The analysis suggests an adequate sample size of 20 participants to offer 80% power. A total number of 20 (13 female; M age = 26, SD age = 6) German-speaking participants took part in Experiment 2a. A total of 21 subjects (16 female; M age = 26, SD age = 5) were recruited for Experiment 2b.

Material

Pseudowords were constructed in analogy to Experiments 1a/b, with the constraint that segment frequencies for onsets, nuclei and coda were controlled for. Frequency calculations (number of words a given segment occurs in) were based on our database (Schmidtke & Conrad, 2018). The final set for valence comprised a total number of 180 pseudowords, with 60 items for each condition of positive, neutral and negative valence (Experiment 2a), comprising 7 different onsets, 3 different nuclei, and a minimum of 3 different codas per condition, (minimum type frequencies: 474, 1375, and 714 respectively, resulting in combinations such as e.g. positive – wiets /viːts/, neutral – fleik /flaɪ̯k/, or negative – drucht /dʁʊxt/).

A total number of 150 items was selected for arousal, containing 50 items for each condition of low, medium and high arousal (Experiment 2b). The reduced sample size as compared to Experiment 1b was attributable to stricter constraints imposed due to higher correlations between arousal SAV and type frequency. Pseudowords were combined based on 5 distinct onsets, 4 nuclei, and 4 different codas per condition (minimum type frequencies: 474, 995, and 760 respectively, resulting in combinations such as e.g. low arousal – dahm /dɑːm/, medium arousal – piet /piːt/, or high arousal – räußt /ʁɔʏst/). For detailed stimulus characteristics, see Table 3.

Table 3. Stimulus characteristics for pseudowords for manipulation of valence (Exp. 2a) and arousal (Exp. 2b).

	SAV’s manipulated for
	Valence (exp. 2a)			Arousal (exp. 2b)
	Negative	Neutral	Positive	Low	Medium	High
SAV Valence	−0.31	−0.05	0.13	0.06	−0.07	−0.25
M (SD)	(0.10)	(0.01)	(0.06)	(0.10)	(0.08)	(0.11)
p	< .001			< .001
SAV Arousal	3.01	2.90	2.89	2.82	2.92	3.03
M (SD)	(0.06)	(0.04)	(0.05)	(0.02)	(0.01)	(0.04)
p	< .001			< .001
Type frequency	3934	3979	4014	3740	3798	3519
M (SD)	(2190)	(793)	(1339)	(1004)	(1524)	(2122)
p	= .96			= .66

Note. M = mean, SD = standard deviation.

Procedure

The procedure was analogous to Experiments 1a/b.

Statistical analysis

The prerequisites of statistical analyses were identical to Experiments 1a/b.

Results

Valence

The fixed factor of valence (with the levels positive, neutral and negative) as well as random intercepts and slopes for participants and random intercepts for stimuli were entered into a cumulative link mixed-effects model. A total of 2.8% of all items were rated as word-like. Excluding all items receiving 2 or more word-like ratings additionally yielded a total amount of excluded items of 13.5%. Our model yielded an effect for the fixed factor of valence, χ²(2) = 10.98, p < .001, R² = .035, again, with more positive ratings for items of neutral SAV (M = 0.458, SE = 0.166) as compared to negative SAV, (M = −0.497, SE = 0.210), z = −3.60, p < .001, Cliff’s delta = .25, 95% CI = [.30, .20] as well as more positive ratings for items of neutral SAV as compared to positive SAV, (M = −0.054, SE = 0.194), z = 2.58, p < .05, Cliff’s delta = .12, 95% CI = [.08, .17]. Ratings for items of negative SAV as compared to positive SAV did not differ, z = −1.92, p = .13 (see Figure 3).

Figure 3. Response patterns for valence ratings for Experiment 2a across different conditions. (a) Estimated response probabilities for each level of valence rating by valence SAV condition (negative, neutral, positive) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each valence condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Arousal

The fixed factor of arousal (with the levels low, medium and high) as well as random intercepts and slopes for participants and random intercepts for stimuli were entered into a cumulative link mixed-effects model. A total of 2.8% of the items were labelled as word-like by participants. Excluding all items receiving 2 or more word-like ratings yielded a total amount of 11.7% to be excluded from further analysis. Our model yielded an effect for the fixed factor of arousal, χ²(2) = 26.88, p < .001, R² = .089, with higher arousal ratings for items in the high arousal SAV condition (M = 0.242, SE = 0.329) as compared to the low arousal SAV condition, (M = −1.449, SE = 0.305), z = −6.28, p < .001, Cliff’s delta = -.37, 95% CI = [-.42, – .32], as well as higher ratings for items in the medium arousal SAV condition (M = 0.042, SE = 0.359) as compared to the low arousal SAV condition, z = −6.06, p < .001, Cliff’s delta = -.32, 95% CI = [-.37, – .27]. Ratings between the levels of medium and high arousal SAV did not differ significantly, z = −0.88, p = .65 (see Figure 4).

Figure 4. Response patterns for arousal ratings for Experiment 2b across different conditions. (a) Estimated response probabilities for each level of arousal rating by arousal SAV condition (low, medium, high) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each arousal condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Discussion

The fact that the effect size of our complete model for valence has decreased when segment frequency has been controlled for, speaks, at least in part, for the validity of the processing fluency hypothesis to be responsible for the more positive ratings of neutral valence SAV. Pairwise comparisons yet still reveal significant deviations from the a priori predicted direction of our effect, as neutral items were continually rated more positive as compared to negative and positive ones also in Experiment 2a, where the confound between SAV and segment familiarity/salience had been controlled for within the material. The fact that negative SAV conditions in both Experiments 1a and 2a received the most negative ratings offers some support for affective iconicity with regard to valence SAV. However, the relationship is not straightforward and needs further clarification.

Concerning the overall models for arousal, in contrast, effect sizes remained largely comparable to those in Experiment 1b when segment frequency was controlled for. Pairwise comparisons reveal the same pattern also in Experiment 2b, following the predicted direction: Arousal ratings were always significantly higher when arousal SAV was elevated as compared to the low arousal condition, regardless of segment frequency.

Taken from a broader perspective, these findings seem to suggest the affective dimension of arousal to be the more relevant driving factor concerning affective iconicity, and a closer look into the internal relations between these two affective dimensions might help to clear the picture of our so far, heterogenous results: As has been shown for word material (see Schmidtke, Schröder et al., 2014; Võ et al., 2009) as well as in the context of pseudoword conditioning (Ando & Kambara, 2023), the two dimensions of valence and arousal are not entirely independent of one another. Rather, a u-shaped quadratic correlation between these two variables suggests low levels of arousal to coincide with neutral valence in our German database (Schmidtke & Conrad, 2018; Schmidtke, Schröder et al., 2014; Võ et al., 2009). However, as valence depicts a bipolar construct – spanning to opposite ends of positive or negative valence from a shared origin – higher levels of arousal tend to co-occur with rising levels of both positive or negative valence. Thus, as suggested by research on (real) word recognition (Kuperman et al., 2014; Snefjella & Kuperman, 2016), differential effects of valence and arousal need to be taken into account.

Considering both the apparently dominant role of arousal as compared to valence and the specific correlation between the two dimensions, an alternative explanation for the counterintuitive finding of more positive valence ratings for neutral as to positive SAV may be found in the fact that neutral valence is generally associated to low arousal, and increasing arousal may have influenced also valence ratings.

To test for this hypothesis, high internal correlations between both variables need to be disentangled to come up with stimulus material that allows testing for independent and potentially differential effects of the two affective dimensions.

Experiment 3a/b and 4a/b

According to the explanations above, pseudowords that entered our experimental conditions were reciprocally controlled for the valence of our arousal – manipulating one dimension while holding constant the other – in the following four experiments. Assuming a possible confound of arousal in Experiments 1/2a, we expected null results for subjects’ ratings of valence for items manipulated on valence with arousal SAV controlled for (Experiment 3a). As a further sanity check, we would additionally expect null results for the same item set when evaluated for any differences in perceived arousal (Experiment 4a)

On the contrary, we expected all effects of arousal to remain stable also under conditions where items were controlled for valence SAV (Experiment 3b). Further, having subjects’ evaluating this same item set in regard to valence, we may nonetheless expect manipulations of arousal to affect perceived valence (Experiment 4b).

Participants

Since the task for Experiments 3a and 4b (valence evaluation task), as well as Experiments 3b and 4a (arousal evaluation task), were identical, we lumped together both item sets from Experiments 3a and 4b as well as 3b and 4a, respectively.

Hence, based on the result of our simulation-based power analysis from Experiment 2b with a reduced stimulus set of 150 items, we had a total number of 22 (16 female; M age = 26, SD age = 5) German-speaking participants take part in Experiments 3a and 4b. Another 20 subjects (12 female; M age = 26, SD age = 6) were recruited for Experiments 3b and 4a.

Material

Pseudowords were constructed in analogy to our previous experiments, with the constraint that valence and arousal were reciprocally controlled for. The final set A for valence comprised a total number of 150 pseudowords, with a reduced size due to stricter constraints based on high correlations between valence and arousal SAV. The resulting pseudowords were comprised of 5 different onsets, 4 nuclei, and 5 different codas per condition (minimum type frequencies: 289, 2204, and 89 respectively, resulting in combinations such as e.g. positive – wehf /veːf/, neutral – rick /ʁɪk/, or negative – trasst /trast/), with 50 items for each condition of positive, neutral and negative valence (Experiments 3/4a). A total number of 135 items (set B) was selected for arousal, using 9, 3, respectively 4 different onsets, nuclei, and codas per condition (minimum type frequencies: 81, 995, and 133 respectively, resulting in combinations such as e.g. low arousal – lohm /loːm/, medium arousal – trach /trax/, or high arousal – spist /ʃpɪst/). Set B contained a total of 45 items for each condition of low, medium and high arousal (Experiment 3/4b). For detailed stimulus characteristics, see Table 4.

Table 4. Stimulus characteristics for pseudowords for manipulation of valence (Exp. 3/4a) and arousal (Exp. 3/4b).

	SAV’s manipulated for
	Valence (exp. 3a/4b)			Arousal (3b/4a)
	Negative	Neutral	Positive	Low	Medium	High
SAV Valence	−0.20	−0.05	0.12	−0.08	−0.09	−1.0
M (SD)	(0.06)	(0.01)	(0.07)	(0.06)	(0.09)	(0.07)
p	< .001			= .59
S AV Arsoual	2.92	2.92	2.91	2.83	2.91	3.0
M (SD)	(0.03)	(0.04)	(0.03)	(0.02)	(0.01)	(0.03)
p	= .41			< .001

Note. M = mean, SD = standard deviation.

Procedure

The procedure was analogous to our previous experiments. The complete set of 284 items (comprised of both sets A and B, with partial overlap between) was presented in one single session that lasted about 20 minutes for ratings of either of the two dimensions valence and arousal.

Statistical analysis

The prerequisites of statistical analyses were identical to Experiments 1a/b.

Results

Valence manipulation (Experiments 3/4a)

The analyses revealed no significant effects for evaluations of valence or arousal ratings (all χ²(2) < 3, see Figures 5 and 6).

Figure 5. Response patterns for valence ratings for Experiment 3a across different conditions. (a) Estimated response probabilities for each level of valence rating by valence SAV condition (negative, neutral, positive) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each valence condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Figure 6. Response patterns for arousal ratings for Experiment 4a across different conditions. (a) Estimated response probabilities for each level of arousal rating by valence SAV condition (negative, neutral, positive) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each valence condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Arousal manipulation (Experiments 3/4b)

Evaluations of Arousal (3b): The fixed factor of arousal (with the levels low, medium and high) as well as random intercepts and slopes for participants and random intercepts for stimuli were entered into a cumulative link mixed-effects model. A total of 3.3% of the items were labelled as word-like by participants. Excluding all items receiving 2 or more word-like ratings yielded a total amount of 15.6% to be excluded from further analysis. Our model yielded an effect for the fixed factor arousal, χ²(2) = 16.23, p < .001, R² = .09, with higher arousal ratings for items in the high arousal condition (M = 0.061, SE = 0.308) as compared to the low arousal condition, (M = −1.597, SE = 0.321), z = −4.89, p < .001, Cliff’s delta = -.37, 95% CI = [-.42, – .32], as well as higher ratings for items in the medium arousal condition (M = −0.288, SE = 0.288) as compared to the low arousal condition, z = −4.08, p < .001, Cliff’s delta = -.31, 95% CI = [-.36, – .25]. Ratings between the levels of medium and high arousal did not differ, z = −1.82, p = .16 (see Figure 7).

Figure 7. Response patterns for arousal ratings for Experiment 3b across different conditions. (a) Estimated response probabilities for each level of arousal rating by arousal SAV condition (low, medium, high) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each arousal condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Evaluations of valence (4b): The fixed factor of arousal (with the levels low, medium and high) as well as random intercepts and slopes for participants and random intercepts for stimuli were entered into a cumulative link mixed-effects model. A total of 6.0% of the items were labelled as word-like by participants. Excluding all items receiving 2 or more word-like ratings yielded a total amount of 31.0% to be excluded from further analysis. Our model yielded an effect for the fixed factor arousal, χ²(2) = 7.55, p < .05, R² = .01. Pairwise comparisons revealed a significant difference for the comparison between items of the medium arousal condition (M = −0.059, SE = 0.150) as compared to high arousing items, with more negative evaluations for the latter ones (M = −0.527, SE = 0.167), z = 2.92, p < .01, Cliff’s delta = .13, 95% CI = [.08, .19]. No other contrasts reached significance (see Figure 8).

Figure 8. Response patterns for valence ratings for Experiment 4b across different conditions. (a) Estimated response probabilities for each level of valence rating by arousal SAV condition (low, medium, high) given in percentages within 95% confidence intervals. (b) Model-based marginal means of post-hoc comparisons for each arousal condition, illustrated with point estimates and error bars representing asymptotic confidence limits given on the logit (not the response) scale.

Discussion

In line with our hypothesis assuming that arousal was the driving force underlying the effects of affective iconicity found so far, no significant differences were found for ratings of valence or arousal following manipulations of valence SAV when items were controlled for arousal SAV. On the contrary, the pattern for effects of arousal SAV on ratings of arousal of pseudowords remained stable, despite controlling for valence SAV. Note, that the overall effect size of our model somewhat decreased as compared to our earlier models, yet still reflects a strong effect. This reduction in effect size may be attributed to the fact that contrasts of manipulations tend to get automatically attenuated when having to control for potential confounds with strongly correlated co-variables.

Most interestingly – even controlling for valence SAV – we even observed a significant effect of the arousal SAV manipulation on valence ratings, this time with ratings being more negative in the condition of high arousal. This particular finding – the perceived valence decreasing with arousal SAV – may account for the counterintuitive aspects of the valence rating data of Experiments 1a and 2a, where neutral items (the ones with lowest arousal) were constantly rated more positive as compared to negative and positive valence conditions that are classically also higher in arousal as compared to neutral valence.

Thus, the findings of the latter four experiments of experimental sets 3 and 4 largely confirm our expectations, indicating that arousal, rather than valence plays a crucial role when evaluating pseudoword material in the absence of semantic meaning solely based on linguistic form.

General Discussion

Previous research has shown sound symbolism beyond motor-sensory features to also extend towards affective dimensions (Adelman et al., 2018; Aryani et al., 2018; Conrad et al., 2022; Körner & Rummer, 2022, 2023; Rummer et al., 2014; Schmidtke & Conrad, 2018). For the present study, following a classical paradigm in sound-symbolic research (Köhler, 1929; Sapir, 1929; Taylor & Taylor, 1965) we asked for subjects’ evaluations of pseudoword material. The series of eight nonword rating experiments in the present study close a gap in the evidence arguing for affective phonological iconicity provided by our research group and others so far:

In several previous studies from our research, we could already show that sublexical affective values influence online language processing – enhancing detection for high arousal sublexical segments in a letter search task (Schmidtke & Conrad, 2018), influencing event-related potentials during lexical decision (Conrad et al., 2022; Ullrich et al., 2016), and modulating lexical access to emotion-laden words (Conrad et al., 2022). Further, SAV has been shown to differ concerning the phonological material used in affectively different types of poems (Aryani et al., 2016; Ullrich et al., 2017).

Taken together, all these findings clearly speak in favour of affective phonological iconicity as being relevant for both the organisation of the lexicon and the processing of language – at least in German – according to the following rationale: As we had computed SAV in a strictly, both elementary and exhaustive, lexical approach – averaging valence and arousal values of all words containing a given phonological segment, if we can find this SAV to produce effects in experimental studies, we can safely infer that (a) sound-to-meaning correspondences in the database are more than mere coincidences, because (b) language users prove sensitive to them.

The missing element in this line of arguments defending the role of affective phonological iconicity, so far, may be seen in the following: How do we know that empirical SAV effects in language processing really involve affective processing? Note that it is generally difficult to clearly separate emotional or affective from attentional processes when interpreting processing advantages for emotion-laden stimuli, both at the level of speeded behavioural responses or ERP effects, because it is well known that (a) emotional content triggers attention allocation and that (b) increased attention facilitates efficiency of cognitive processes. Clearly, the ERP-study of Conrad et al. (2022) provides additional important evidence pointing towards an internal relation between the signifier and the signified – a crucial aspect of iconicity – showing that congruence between sublexical and lexical affective values enhances lexical access. However, as the resulting N400 ERP effect only represents the final integration result of a whole series of perceptive, or cognitive processes of different grain size, from visual features to semantics – including articulatory processes even in a silent reading task, it remains debateable how exactly this congruency – although clearly operationalised – would involve purely affective processes at the sublexical level. Thus, it appears challenging to draw precise conclusions from the data of Conrad et al. (2022) on how exactly the integration between sublexical phonology and affective semantic meaning of words gives rise to the affective iconicity effect present in their data, and to what extent affect (as compared to e.g. attentional or articulatory processes) is involved in the processing of sublexical phonology – considering the composite nature of the reported ERP data.

On the other hand, the study of Aryani et al. (2018), convincingly shows how acoustic sublexical features elicit affective impressions, but the experimental part of this study does not allow for conclusions concerning affective iconicity in the vocabulary of German, as the definition of sublexical affectivity in that study was – at least in important parts – no longer related to the occurrence of segments across the lexical affective space (but instead to ratings of how arousing words or pseudowords sound – regardless of the semantic meaning).

Further, most empirical studies available to date on affective iconicity do not allow for differentiating between effects of arousal versus valence as both dimensions are highly correlated in natural stimuli. Our innovative design, manipulating respective dimensions separately in all experiments, and explicitly disentangling the arousal and valence SAV in Experiments 3 and 4, aimed to overcome this theoretically important shortcoming.

In sum, the present study closes these gaps in that we can show how sublexical affective values derived from our normative lexical database of valence and arousal values for 5695 German words, influence the affective impressions evoked by otherwise semantically meaningless pseudowords. In all of our experiments, pseudowords constructed via permutation of syllabic onsets, nuclei, and codas, that have a tendency to occur more often in words of medium or high arousal – as compared to low arousal – are perceived as more arousing compared to pseudowords comprising segments occurring mostly in words with low arousal semantic meaning, receiving affective “meaning” via SAV.

Note that the same, straightforward, evidence of affective phonological iconicity is not given in our data concerning valence: Although Experiments 1a and 2a feature some evidence for negative SAV segments being evaluated more negatively than neutral or positive segments, this apparently congruent result stands in contrast to the surprising finding that, in both experiments, neutral conditions received more positive evaluations than positive ones.

The reduction of respective effect sizes following the control of formal salience, or segment frequencies, in Experiment 2a, speaks somehow in favour of familiarity effects – according to the processing fluency hypothesis (Belke et al., 2010; Reber et al., 2004) evoking more positive affective impressions for more frequent sublexical segments. Note that SAV of more familiar, more frequent segments naturally tend to be more neutral simply because, the more words a given segment occurs in, the more likely the average of valence and arousal values for these words approaches 0, the neutral centre of the valence scale.

Controlling for arousal SAV in Experiment 3a however, revealed that all apparent valence SAV effects – including the more negative ratings for negative valence in Experiments 1a and 2a – vanished. Furthermore, high arousal of phonological segments, even evoked more negative valence impressions in Experiment 4b despite the fact that valence SAV had been controlled for.

While our method relies on visual presentation, we are confident to attribute our empirical effects to phonological rather than orthographic processing, because our manipulations relate to the phonologically defined sublexical units of syllabic onsets, nuclei and codas, and the mechanisms of automatic phonological and prosodic recoding of written words during silent reading are very well-documented in visual word recognition (see Jacobs & Grainger, 1994 for review; and Jacobs et al., 1998, for computational model): Over the past twenty years, a wealth of research in visual word recognition, including behavioural, computational, and neuroimaging studies, has consistently shown that phonological information is automatically and involuntarily produced from written words, playing a crucial role in facilitating lexical access, as demonstrated by various studies (for example, Braun et al., 2009; Conrad et al., 2007, 2009, 2010; Ziegler et al., 2000, 2001; Ziegler & Jacobs, 1995). Furthermore, as phonological units are represented in a dual manner, encompassing both auditory and motor aspects as outlined by Hickok (2012), neuroimaging research has shown that the motor circuits responsible for articulatory movements are activated when word stimuli are presented visually (see e.g. Burton et al., 2005; Hagoort et al., 1999), suggesting articulatory activations, particularly in relation to complex phonological clusters.

Thus, as mentioned in the introduction, the highly transparent grapheme-phoneme conversion rules of German make this language well suited for our approach to affective phonological iconicity even when using visually presented stimuli.

Taken together, all these findings sum up to a clear-cut picture:

Affective arousal – as a basic constituent of semantic meaning (see Osgood et al., 1957 for the seminal semantic differential study; see also Schauenburg et al., 2019, for respective effects in sentence processing) is marked at a sublexical phonological level in systematic ways in the vocabulary of German, and the systematicity and importance of this apparent affective iconicity in the database is corroborated by the empirical finding that the underlying sound-to-meaning correspondences influence the affective impressions of language users – even when presented in otherwise “meaningless” pseudowords. Therefore, our data suggest that these sound-to-meaning correspondences do possess “psychological reality”.

Further, these arousal effects clearly override valence effects in our data, therefore we conclude that arousal rather than valence is the relevant dimension concerning affective iconicity. These novel data importantly add to previous proposals of effects for salient phonological units in the range of negative, high arousal affect, consistent with general principles of behavioural adaptation (e.g. Adelman et al., 2018). Our specific experimental design where arousal was manipulated while valence was controlled for, and vice versa, enabled us to provide a more differential view in favour of arousal alone being responsible for these iconicity effects.

This distinction between the two dimensions seems to make perfect sense taking into account more general proposals of emotion theory (Schachter & Singer, 1962; see e.g. Scherer & Moors, 2019, for more recent reiterations): Whereas arousal is often understood as a measure of emotion intensity, it can also be seen as representing an initial fast alert process, enabling the organism to respond quickly to potentially emotionally relevant stimuli. Valence, on the other hand, would be more relevant for more conscious evaluation processes of the stimuli – determining whether a stimulus initially inducing an alerted state, is good or bad (see also Imbir et al., 2023, for a discussion of the nuanced interplay between arousal and valence with possible contextual effects depending on subjective significance).

Note that ERP studies – particularly sensitive to the time course of brain processes – offer support for this view of early arousal effects preceding those of valence: In Recio et al. (2014), for the presentation of visual word stimuli, in an experimental design orthogonally crossing the experimental factors of semantic valence and arousal, we found evidence for pure arousal effects – regardless of valence – starting at 200 ms, whereas valence effects, when arousal was controlled for, did not influence ERPs before 275 ms after stimulus onset (see also Sulpizio et al., 2021). Both concerning these ERP data and the nonword rating data of the present study, we, therefore conclude that this pattern of results is in accordance with the view of arousal as a valence-unspecific alerting system that operates at the early stages of processing by triggering the allocation of attentional resources to prepare the organism for an evaluation of events in terms of pleasant/unpleasant and further emotional responses (e.g. Bradley & Lang, 1994; Herbert et al., 2006).

For sublexical phonological units, the processing of which clearly occurs early during language processing – before semantic information becomes available – it seems, therefore, conclusive that the kind of affective information the language system refers to when marking a given word at a sublexical level as especially relevant, because of its affective semantic content, would rather be arousal, and not necessarily valence. The more elaborate evaluation processes corresponding to the processing of valence would occur at later processing stages where semantic information has become available. Our data reveal how during these conscious evaluation processes – in the case of the ratings our participants had to deliver – increasing arousal even led to more negative evaluations. Despite the fact that valence SAV of our stimuli in Experiment 3a and 4a were controlled for, no effects for valence manipulations alone were any longer observable.

In conclusion, our study provides compelling evidence for the significance of affective phonological iconicity in the lexicon, particularly emphasising the role of arousal over valence in shaping linguistic perception and processing. Our results suggest affective properties embedded at a sublexical level of language to carry measurable impact on how greater linguistic units as pseudowords or words are perceived and processed, underscoring the intricate ways in which affective information is encoded and decoded in language.

While at a theoretical level, our findings hold implications for theories of language evolution, suggesting a potential role for emotional salience in the development of phonological systems across languages, we consider that this knowledge could also be beneficial whenever newly to be created language shall be made most efficient combining a “higher level” affective semantic meaning with consistent sublexical phonology, e.g. designing brand or product names in commercial contexts or formulating messages in social or political contexts. Further, as the rationale of our computations is easily reproducible, we look forward to potential replications of our findings in other languages – exploring cross-linguistic aspects of affective iconicity. To that extent, see, for instance, a preprint of Gatti et al. (2023), who, using distributional semantic models, were able to predict pseudowords’ valence based on letters and bigrams (i.e. objective components) in English, or Calvillo-Torres, Haro, Poch, Ferre, & Hinojosa, (submitted manuscript) for a report on systematic distribution patters of phonemes with regard to valence and arousal in the Spanish vocabulary, for very recent additional support on non-arbitrary relations between affect and formal aspects of language in other Indo-European languages.

In sum, our data provide a straightforward cross-validation account linking corpus-linguistic and experimental approaches to the intriguing documentation of arousal-related affective phonological iconicity marking word forms at a sublexical phonological level.

Acknowledgements

David Schmidtke: investigation, formal analysis, writing – original draft. Markus Conrad: conceptualisation, investigation, supervision, writing – original draft.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported by a grant to M.C. (410 “Sound physiognomy in language organisation, processing, and production”) from the German Research Foundation (DFG) via the Cluster of Excellence “Languages of Emotion” at the Free University of Berlin.