Antonym adjective pairs and prosodic iconicity: evidence from letter replications in an English blogger corpus

1.1 Iconicity

It is frequently assumed that one “design feature” of language is that there is no relation between sound and meaning (Hockett 1960), which is what makes languages so diverse and flexible. However, there is accumulating empirical evidence that this assumption is only partially true, and that some degree of iconicity, i.e. an association between form and sign, may have been necessary in certain periods of humankind and human interaction. Perniss and Vigliocco (2014) argue that “iconicity is a fundamental property of language, representing an adaptation to a critical constraint on the phylogenesis, ontogenesis and use of language, namely the need to map linguistic form to human (sensory, motor and affective) experience” (Perniss and Vigliocco 2014: 2).

One well-known example of iconicity involves systematic shape-sound correspondences such as the so-called “bouba-kiki” effect, where rounded and non-rounded mouth shapes and tongue movements in production are systematically associated with rounded and pointy shapes, respectively (Köhler 1929; Ramachandran and Hubbard 2001; Maurer et al. 2006). As evidence of the widespread nature of iconicity, Blasi et al. (2016) analyzed 100 basic vocabulary items in 4,298 different languages and 359 lineages, finding a large proportion to exhibit non-arbitrary sound-meaning relations that cannot be explained on the basis of language contact. For example, words for “tongue” tend to contain /l/, while those for “nose” tend to contain /n/. Thus, iconicity in spoken language may be less exceptional than expected. It has even been studied in the context of consumer judgements in brand names (Yorkston and Menon 2004): brand names for ice cream differing only in the two phonemes /i/ and /ɛ:/ (or /ʌ/^[1]) revealed different consumer judgements on richness, smoothness, and creaminess. Brand names with /ɛ:/ were associated with heavier, richer, and creamier ice cream. This association is based on sound symbolism, specifically the tendency to associate front vowels with “smallness, lightness, mildness, thinness, fastness, coldness, bitterness, femininity, weakness, lightness, and prettiness” (Yorkston and Menon 2004: 44) in comparison to back vowels. The authors suggest that the preference for one or another product based on sound symbolism is below the customer’s awareness.

1.2 Adjectives and scalar meaning as a semantic domain for iconicity

A semantic domain where iconicity is well documented involves scalar dimensions such as size, duration and speed. Such meanings are expressed by words of a variety of grammatical categories, including nouns, verbs and prepositions, but most typically represent the semantic content of gradable adjectives such as large/small, long/short and high/low. Gradable adjectives as a class can be identified by their occurrence in the comparative form (e.g. larger, longer). Formally, they may be analyzed as lexicalizing mappings from individuals to degrees on scales (Bierwisch 1987). Typically, such adjectives occur in antonym pairs, where one adjective is associated with the “large” direction on the scale while its antonym is associated with the “small” direction. Often, the adjective that corresponds to the “large” direction is the unmarked one. This manifests itself in several ways, one being the interpretation of questions: “How long is X?” does not presuppose that X is long, whereas “How short is X?” presupposes that X is short. However, this is not always the case. There are adjectives denoting very high degrees, like huge, that are marked, and there are antonym pairs where neither member is unmarked (e.g., poor and rich). Below we will see a further instance of divergence between largeness and unmarkedness that will prove relevant for the understanding of iconic effects in this domain.

A variety of factors make scalar meaning a natural place to expect iconicity in language, including humans’ ability to associate magnitudes across modalities (Cuskley and Kirby 2013) as well as regular co-occurrences in the external world, such as that between the size of animals and the fundamental frequency and resonance of their vocalizations (Ohala 1984, 1994, 1997). Larger animals produce lower vocalizations because their vocal folds tend to be longer and heavier, resulting in a lower voice in comparison to smaller ones (Ohala 1994, citing Morton 1977). Hence, based on the acoustic properties of the vocalizations, animals can roughly estimate the size of a potential aggressor and behave in the appropriate manner.

Recently, Knoeferle et al. (2017) tested the role of the acoustic properties in humans. They presented participants with visual representations of objects of different size, accompanied by audible nonsense speech material that differed with respect to vowel duration, formants, intensity and fundamental frequency. Participants had to judge which object would be a good match to the audible sound. Their results revealed an iconic effect of object size on vowel duration and the first formant (F1), but not in the other parameters. Larger objects were associated with longer duration and higher F1 values.

1.3 Produced and perceived iconicity in the prosody of adjectives

Prosody can also contribute to iconicity in spoken language, for example by an extra lengthening of the vowel in long in referring to the experience of an event taking a long time. Schlenker (2018) has recently proposed that such lengthening is an iconic enrichment of the arbitrary encoding of linguistic meaning. Iconic prosodic modulation consists of changing certain prosodic features such as duration, fundamental frequency (F0), or amplitude to express additional meaning. Along similar lines, Nygaard et al. (2009) investigated the relationship between prosody and meaning in two experiments. Three speakers had to read novel words in an infant-directed speech style to stimulate a situation of engagement. The novel words, embedded in a frame sentence, were first read with a relatively neutral prosody as a baseline. Subsequently, the words were presented together with pictures that represented the meaning of an adjective from the pairs happy/sad, hot/cold, big/small, tall/short, yummy/yucky, and strong/weak. Speakers had to read the words again. Differences in acoustic properties were found (mean F0, F0 variation, duration, and amplitude) depending on the adjective. The authors also investigated whether listeners could reliably infer the meaning of these novel words. For this purpose, listeners saw two pictures representing an antonym pair, heard one of the previously recorded sentences, and had to choose the picture that would correspond to the perceived novel word. Listeners were significantly better in choosing the right picture when listening to the speaker’s meaningful prosody than they were when hearing neutral prosody, as well as with the matching prosody rather than the mismatched prosody. The authors suggest that prosody could augment, disambiguate, or reinforce meaning.

Perlman (2010) showed video clips involving fast and slow events to experimental participants, who were asked to retell the different events; they did so by generally talking faster for the faster events and slower for the slower ones, without being instructed to do so. In a more recent study, Perlman et al. (2015) extended these findings from the manner of motion to the size of an entity and additionally from concrete to abstract meanings (e.g. concrete: a fast drive; abstract: slow career progress). Speakers had to read short stories involving one of these semantic dimensions to a partner. The authors predicted that stories with different manners of motion would elicit corresponding prosodic variation in duration, but not fundamental frequency, while the reverse would be the case for stories varying adjectives corresponding to the dimension of size. Their findings show that stories in the small condition were read with higher F0 than stories in the large condition, for both abstract and concrete meanings. Moreover, stories in the fast condition were read within a shorter duration than stories in the slow condition, and no differences in F0 were found. Thus, different acoustic parameters may be used to mark different semantic dimensions.

In three experiments, Shintel et al. (2006) recorded speakers who described either a dot moving in upward and downward directions or dots moving with different speeds. They showed that motions in the vertical dimension go hand in hand with changes in F0 in a similar direction, whereas changes in speed coincided with changes in speech rate. In follow-up studies, Shintel and Nusbaum (2008) also provided evidence that the speed of recorded instructions influenced the time of a listener’s response. In Shintel et al. (2014) the findings were extended to novel-word learning, showing that congruent prosody has a positive effect on memory consolidation.

The iconic representation of scalar meaning is not limited to the spoken modality. In signed languages including American Sign Language (ASL) and Italian Sign Language (LIS), adjectival and verbal scales are in some cases visually represented in the signing space, a pattern that has been characterized as iconic (Wilbur et al. 2012; Kuhn 2017; Aristodemo and Geraci 2018). In LIS, for example, the comparative taller can be signed via hand movement in the upward direction. Furthermore, the production of signs may be modulated with semantic effect, as when the slower-than-normal signing of a verb conveys that the corresponding event was a slow one (Wilbur 2008). Similar effects are discussed in Schlenker et al. 2013. In ASL, the sign GROW can be realized with different speed and different maximal distance of the hands. In the ASL translation of the sentence My group has been growing, the named parameters – speed and amplitude – are decisive for the interpretation of the sentence. Depending on these parameters and the iconic mapping, it can express that the group grew quickly or slowly, and that it grew a lot or only a bit.

Here, the question naturally arises as to what exactly should be counted as an ASL sign and what is actually gesture (see Goldin-Meadow and Brentari 2017 for an overview and discussion). Are all these individual manipulations on the grammatical side, belonging to the sign as such, or are they merely gestural modifications of the sign? In spoken language, speech-accompanying gestures can take over this part of modifying what is said in iconic ways (see Kendon 1980; McNeill 1992 for discussion of speech-accompanying gestures and their iconic character). For example, when talking about a painting, the utterance can be accompanied by an oval gesture or a rectangular gesture, thus indicating whether the painting is actually oval or rectangular.

Schlenker (2017) distinguishes between external and internal (i.e. syntactically (in)eliminable) enrichments. Speech-accompanying gestures in spoken language would be of the former kind, while the prosodic modulations we discussed in the beginning would be of the latter kind, as would the above-discussed modulations of speed and amplitude in sign languages.

1.4 New prospects on prosody with recent technological developments

Along with progress in technology (e.g. computers, smartphones, tablets, fiber optic cables, satellites), digital communication has come to have an enormous impact on our daily life and our communication tools and styles. Social media platforms (e.g. Twitter, Instagram, Facebook, chats, WhatsApp, blogs) have been developing in parallel.

Social media data have some features in common with spoken language, since writers do not follow all the formal rules of traditional written norms, and their utterances are further enriched with icons of emotional expressions (emojis). These platforms provide a great opportunity to investigate the dynamics and creativity in the use of written language beyond prescriptive rules (Huang et al. 2016; Kaye et al. 2017). Using social network databases provides the advantage of getting a vast amount of data in completely natural settings, with participants who do not feel constrained as they might in a laboratory experiment. The disadvantage, however, is that the data might be confounded by many unknown factors. So far, most studies have tested for iconicity in prosody using contrastive settings instead. Perlman and colleagues write that “Remarkably little is known about how speakers use iconic prosody in the wild” (Perlman et al. 2015: 1349). Our work contributes to this research by using a social media corpus.

Among others, the idea of using social media for a better understanding of prosody goes back to Brody and Diakopoulos (2011), who considered word lengthening by letter replications as a substitute for prosodic emphasis. These letter replications are a way of signaling the writer’s sentiment and emotion in written text where some properties of spoken language, such as intonation, are partially absent. Letter replications also represent a possible candidate for a feature of written language that may have an iconic effect, in that the lengthened pronunciation of a word such as long in spoken speech can be reflected in writing via replications, as in looonng. To date, however, this has not been systematically examined for antonym pairs.

With this work, we argue that lengthening a word (as in looonng) has an iconic effect. In other words, a speaker or the author of a written text uses word lengthening to express some iconic feature of the object, property, or event under discussion (here: that something was extremely long). We do not, however, intend to exclude the possibility that there might be other reasons for word lengthening in spoken and written language as well; in particular, this strategy might be used to express amplification or emphasis (see Kawahara and Braver 2014; Fuchs et al. 2018).

1.5 Research questions and expectations

In this paper, we investigate the presence of iconic prosody in written language. Based on the literature on prosody and iconicity, we focus on scalar meaning as expressed by gradable adjectives, and on letter replications as a prosodic feature with a potentially iconic effect. The following specific research questions were asked:

1. In which adjectives does lengthening (letter replications) occur?

   Here we distinguish two hypotheses. H1) If, as we assume, prosody and letter replications have some degree of iconicity, we predict that letter replications will occur more frequently for adjectives that express the larger size, e.g., in long rather than in short. H2) If, on the other hand, letter replications are used exclusively as an expression of emphasis or prosodic amplification, we predict a comparable frequency of letter replications in both members of a pair of antonyms.

2. If replications occur, how many letters do they comprise?

   Consistent with the view that letter replications have an iconic component, we expect longer words (i.e., a greater number of replicated letters) for the larger size dimension.

2.1 The blogger corpus

An English social media corpus was used. The corpus is freely available for non-commercial use. It consists of approximately 140 million words written by 19,320 bloggers in August 2004 (Schler et al. 2006). The age of the bloggers ranges from 13 to 47 (in three age groups) with an equal number of males and females. The profession of the bloggers is also provided. In the present paper we do not investigate potential age and/or gender effects in the frequency of letter replications with gradable adjectives; we refer the reader to Fuchs et al. 2018 for evidence that, in other domains, it is younger bloggers who make the most use of letter replications.

2.2 Selected adjectival antonym pairs

Our starting point for selecting antonym pairs for investigation was the literature on iconic prosody in spoken language discussed in the introduction. In particular, we took from Shintel et al. (2014) all those adjective pairs with a dimensional meaning (i.e., a meaning relating to size or spatial/temporal extent). In order to ensure that our data set was sufficiently comprehensive, we augmented this set of pairs by extracting all words tagged as adjectives from the list of the top 5000 words/lemmas in the Corpus of Contemporary American English (Davies 2008),^[2] and further restricting this list to those adjectives judged to have a dimensional meaning. The resulting complete set of adjectives was organized into antonym pairs. In some cases, this procedure resulted in two or more adjectives being associated with a single antonym; for example, long and tall share the antonym short, while huge, giant and enormous share the antonym tiny. In order to arrive at distinct pairs for analysis and avoid double counting of certain adjectives, we selected in each case one adjective for the “large” end of the scale and one adjective for the “small” end. In most cases where a choice had to be made, the adjective selected was the one occurring most frequently in the blogger corpus used in the present study (e.g., long was selected for pairing with short because it was more frequent than tall), with the following exceptions: i) little was excluded as a possible antonym for big because it has a quantificational and degree modifier use in addition to the adjectival one; ii) close was excluded as an antonym for far because it also has a use as a verb; iii) young rather than new was selected as the antonym for old for consistency with Shintel et al. 2014. The final list of antonym pairs is shown in Table 1.

In each case, one member of the pair expresses a greater degree of size, extent or duration, while the other expresses a lesser degree. Here we note that the ordering of the pairs in Table 1 is based on “largeness”, not markedness. As discussed in the introduction, these properties typically coincide. However, the pair slow and fast represents an exception: fast is arguably the unmarked term (cf. “How fast is the train?” vs. “How slow is the train?”), but it is instead slow that is associated with larger temporal extents, because a slow event requires more time than a fast one. We thus expect slow to be targeted for lengthening more often than fast.

Note also that the adjective pairs short/long, near/far, fast/slow, narrow/wide, thin/fat and young/old correspond to the horizontal axis, whereas low/high and shallow/deep correspond to the vertical axis and small/big and tiny/huge do not make specific reference to the axis but are instead general size properties.

2.3 Data extraction and preprocessing

The NLTK toolkit was used as a natural language processing environment to tokenize the corpus (http://www.nltk.org/api/nltk.tokenize.html). All lower- and upper-case tokens were considered together. To further process the data, we used Python 2.7. and R (R Core Team 2017).

In a first step, all replications of letters were extracted and removed from the selected adjectives (cf. Table 1). This resulted in strings corresponding to a specific order of successive letters for each word and also included replications which are the orthographic norm (e.g., double l in small was removed, yielding smal). From the resulting corpus the following words were eliminated: all words that do not involve all the letters of the original word (e.g. narow instead of narrow), all words that differed in just one letter and could potentially be typos (e.g. thinn instead of thin), and, finally, all words that might have a different meaning (e.g. tinny instead of tiny).

Furthermore, we calculated the overall number of cases for each adjective with and without letter replication. The number of cases without letter replications served as a baseline indicating how often bloggers wrote the specific word in the orthographic norm. The total number of cases including letter replications was set to 100 percent for each adjective and then subtracted by the baseline in percent. The result corresponds to the percentage at which bloggers wrote the selected adjectives with letter replications.

In addition, we calculated the length of each word as the number of letters in the orthographic norm and the length of each word including replicated letters.

Figure 1:

Plots with percentage of occurrence with letter replications in the respective adjectives with respect to the baseline (y-axis). Data are split by adjective group (adjectives expressing larger degrees on the left and smaller degrees on the right).

Figure 2:

Stacked bar plots with replication length (number of letters for words with replications only). The number of letters for the default word was subtracted from the overall number of letters to account for differences in word length among adjective pairs). Adjective pairs are written on the x-axis. All data are given in percent. Percentages in different colors are associated with different categories.^[3] Category 2, for instance, is a word with two more letters, category 5 with five more letters and so on.

3.1 Percentage of adjectives with letter replications

Almost all antonym pairs show a numerically higher frequency of letter replications in the adjective corresponding to the “larger” scalar direction (cf. Table 2). Letter replications are used in the adjectives long, slow, big and huge significantly more often than in their respective antonyms. For example, out of all occurrences of long in the corpus (n = 44819), in 1.35 % of the cases the word was spelled with letter replications, i.e. different from the orthographic norm. In the case of far and old, the difference falls short of being significant. But importantly, in no case was the frequency of replications higher for the adjective corresponding to the “smaller” scalar dimension. Thus, overall we find support for our first hypothesis H1 rather than the second hypothesis H2. We note, however, that the effect was not found consistently across all pairs tested, but instead was strongest in pairs relating to the dimensions of overall size (small/big, tiny/huge) as well as temporal extent and spatial extent in the horizontal direction (long/short, fast/slow), and not evident for extent in the vertical direction (low/high). We return to this point in the conclusion section.

The percentage of occurrences of letter replications can be seen in highly frequent words in the corpus (e.g., long with n = 44,819), but also in less frequent words (e.g., in slow with n = 4,696). The group of adjectives expressing larger degrees is, however, realized much more frequently than the group of adjectives expressing smaller degrees (cf. Table 2).

Figure 1 summarizes these findings by comparing the two adjective groups. The group of adjectives expressing larger degrees are not only found more often, they are also more variable than the group of adjectives expressing smaller degrees.

3.2 Number of repeated letters in words with letter replications

In a next step, we investigated whether the number of repeated letters (in words with letter replications) would also differ among the adjectival antonym pairs (cf. Appendix). For this purpose, we subtracted the number of letters in words with replications from the number of letters in the orthographic norm. Furthermore, we categorized all words with letter replications in categories from 2 to 10+. For example, category 2 would correspond to words with two additional letters, category 5 to words with five additional letters and category 10+ to words with ten or more additional letters. The stacked bars in Figure 2 give an overview of which categories are realized more often than others. Adjectives where we found no letter replications are not included in the plot. Results are depicted in percent. In contrast to our expectations, no consistent differences between adjectives with larger and smaller degrees are visible. In most adjectives, words are replicated with three or four additional letters.

Note also that the stacked bars corresponding to adjectives with smaller degrees include fewer data points than adjectives with larger degrees. On the basis of these findings we can exclude the possibility that writers on average lengthen all words for larger degrees to a greater extent than words for smaller degrees. Antonym pairs were, however, not obtained in a contrastive context, e.g. fasssttt versus sloooooooooooow. It might well be possible that differences occur when the adjectives are used as antonym pairs.