Microsoft Word - hunter-smith

Understanding without Babblefish: Reviewing the Evidence for Universal Sound

Symbolism in Natural Languages

Sarah Hunter-Smith

Swarthmore College

Abstract

Sound symbolism refers to a non-arbitrary relationship between

sound and meaning. Language-specific forms of sound symbolism are

well documented, but many scholars have also been interested in whether

some sound symbolic patterns are universal. If humans have common

intuitions about how sound should represent meaning, these intuitions

could have facilitated the origin of language. If humans share sound

symbolic intuitions, and they were influential in the origin of language,

then we would expect to see evidence of these patterns in natural

languages. Some evidence has been found in size ablauting systems,

deictic pronouns, and ethnozoological nomenclature, however, one study

in particular, Brown et al. (1955), suggests that sound symbolic patterns

may be evident even among sensory adjectives. Subjects in this study

were able to correctly match a pair of sensory antonyms in a foreign

language to their English translations at rates significantly above chance.

By conducting a similar study using a well-described sound symbolic

pattern to create a “symbolic” and “non-symbolic” list of the word pairs, I

show that subjects need symbolic cues to perform at levels above chance.

I further try to show that this was likely true of the subjects in the Brown

et al. study as well. This suggests that sensory adjectives reflect human

intuitions about how sound should represent meaning. A rough assay to

determine the extent to which big-small in many languages conforms to a

sound symbolic pattern, however, failed to find the pattern represented

more often than would be expected by chance. This suggests that sound

symbolic intuitions place at most a subtle constraint on sensory adjectives.

In general, sound symbolic patterns seem to be represented inconsistently

in language, perhaps because there is an opposing selective advantage of

arbitrariness in language.

1 Introduction

Sound symbolism is generally defined as a direct connection between sound and

meaning. This connection may be imitative, as when we use boom to mimic the sound

of an explosion. Alternatively it can be metaphoric, as one would find in an explanation

I would like to gratefully acknowledge my thesis advisor, T. Fernald, and my second faculty reader K.D.

Harrison for their helpful comments and guidance. I would also like to thank T. Johnson and B. Woods

for proofreading and offering insightful comments. A special thanks to S. Byrd-Krueger and M.K.

McInerney for their help distributing the survey, and to A. Gillis and N. Peters for their assistance in its

design. Finally I would like to express my gratitude to the 56 people who took the time to participate in

the survey, and also to the many friends who have patiently listened and acted as a sounding board for

the last three months.

for sound patterns used in a poem. For example, we might say that the frequent use of

words containing voiced stops in a certain poem evokes a sense of heaviness or power.

Most people have at one time or another expressed the idea that the way a certain word

sounds is particularly apt given its meaning, so the idea that certain sounds have certain

meanings is not foreign. At the same time, this notion is in conflict with a basic

observation about human language. Not all humans speak the same language. The

group of sounds which mean dog in English are different and unrecognizable as having

any relation to the group of sounds with the same meaning in Japanese. A general tenet

of linguistics, therefore, has been that the sounds composing a word are nearly always

arbitrary and unrelated to that word’s meaning (Fromkin et al. 2003). This leaves us

with some questions. Is onomatopoeia just a small exception to this principle? Do our

strong feelings about what sounds are heavy have any psychological reality for other

people, even people who speak different languages?

At the root of this conflict is an issue that has fascinated many scholars: whether

sound symbolism is only language specific or whether it also has universal elements.

Sound symbolic sub-systems within specific languages are a well supported

phenomenon at this point (Childs 1994, Austerlitz 1994, Hamano 1998, Blust 2003). Not

only can we find sound-meaning correlations, but most speakers exhibit statistically

significant intuitions about these correlations in their language (Parault and

Schwanenflugel 2006). Universals of sound symbolism on the other hand seem to be a

bit like the Holy Grail: exciting but somewhat speculative. Language-specific sound

symbolism can teach us interesting things about how humans actually process

language, but if no elements of it are universal, then it is ultimately conventional and

arbitrary just like other aspects of language. Universal sound symbolism is exciting

because it raises the possibility that we as humans are predisposed to connect certain

sounds with certain meanings. This would have greatly facilitated the evolution of

language, and could give us important insights on this notoriously difficult puzzle. Yet

there seems to be considerable counter evidence to universal sound symbolism and

there is no conclusive answer yet as to whether it actually exists.

A fascinating experiment: Brown et al. (1955)

One attempt to address the question of the existence of universal sound

symbolism was an experiment done in 1955 by Brown et al. This work looked for

evidence of universal sound symbolism by examining whether people could to some

extent assign the correct meaning to foreign words. The authors proposed that if

humans are predisposed to represent concept A with sounds of type B, then this should

be reflected in natural languages, and furthermore humans should connect sounds of

type B with concept A when they encounter them. Clearly people cannot translate

unfamiliar foreign words when presented with them, but given a pair of antonyms

which describe two ends of a sensory spectrum and the translation of these words in a

foreign language, subjects might be able to say that one of the foreign words sounds

more like one end of that spectrum than the other. Brown et al. presented subjects with

twenty-one such pairs in English and their translations in Chinese, Czech, and Hindi.

Surprisingly, subjects correctly matched the pairs on average at frequencies

significantly above chance for all three languages. This was true whether they read the

words and heard them pronounced by a native speaker, or just read the words silently.

This controlled for the possibility that the person reading the words aloud might use a

certain intonation or tone that could give the subjects clues about a word’s meaning.

Subjects in this study did not do equally well for all pairs, however. The percentage of

subjects that guessed correctly on a given pair ranged from 17% to 97% (the results are

presented in Appendix II). In other words, subjects agreed strongly about the way

some word pairs should be translated and were correct, agreed strongly about other

word pairs and were wrong, and also had mixed judgments about some pairs, but

overall they were able to guess the correct translations more often than not.

This study raises a number of interesting questions which have not been

convincingly addressed in the intervening years:

1) Were subjects able to correctly match word pairs based on some sound symbolic

quality? Brown et al. determined that differences in word length did not

predict subjects’ responses, but it is conceivable that a variety of random

factors like word length could have been responsible. For example, if words

had superficial resemblances to their English translations, this could explain

why subjects agreed strongly in some of their responses.

2) Why did subjects perform much better on some word pairs than others? If this

result was not due to random patterns in the data, it could indicate that

certain semantic domains exhibit sound symbolism more widely than others.

Identifying these semantic domains could give us clues about the

neurological basis of universal sound symbolism.

3) Would subjects do as well on a wider sample of languages? The results suggest

both that subjects have intuitions about how meaning should be represented

by sound, and that their intuitions are reasonably well reflected by patterns in

natural languages. If these patterns are in fact widely distributed, then there

might be some basis for the hypothesis that sound symbolism played a key

role in the origin of human language.

This paper addresses these three questions as a basis for evaluating the existence

and possible nature of universal sound symbolism. I show that subjects do use sound

qualities to make judgments in an experiment similar to Brown et al. by comparing

performance on a group of word pairs that contrast a particular sound quality and a

group that does not. Further analysis of the data in Brown et al. (1955) suggests that

subjects in that experiment also predominantly used sound quality over superficial

factors. It also suggests that subjects most successfully recognize word pairs which

represent basic sensory parameters such as brightness/darkness or

sharpness/bluntness. However, despite the success that subjects seem to have using

sound quality, attempts to quantify a relationship between sound and meaning in

natural language have generally failed to find significant correlations, and they fail in

this paper as well. Universal sound symbolism may exist in the form of shared

intuitions about sound-meaning associations, although this has yet to be adequately

demonstrated. The evidence suggests, however, that these shared intuitions are not

consistently reflected in natural languages. I conclude by discussing the limited

distribution of sound symbolic patterns in natural language in the context of the

competing selective advantages of incorporating sound symbolism and arbitrariness

into language.

2 Background

2.1 Types of Sound Symbolism

Before continuing, it is necessary to define sound symbolism more carefully. The

term sound symbolism may refer to as many as four related phenomena which are

differentiated by Hinton et al. (1994). The explanations below are adapted from this

source, but include some additional examples and comments.

Corporeal sound symbolism. This describes sounds outside of language which are

directly related to meaning by indicating the emotional or physical state of the speaker.

Physical state can be indicated by sounds such as coughing or hiccupping. Emotional

state can be indicated by pitch level, range and variability, loudness, and tempo

(Ostwald 1994). Corporeal sound symbolism has many universal elements and some

aspects are even shared with other animals. This kind of sound symbolism is of little

interest in this paper apart from the fact that it may share a biological origin with other

kinds of symbolism.

Imitative sound symbolism. This category includes all sounds which are imitative

of other sounds whether accurate or conventionalized. That is, it includes both

someone’s direct mimicry of the sound their cat makes, and also the word meow. Purely

imitative sounds should be universally understood, provided that the source of the

sound is widely distributed, but conventional sounds may be different from language

to language. However, given their basis in imitation, we might expect that these

conventional sounds would be more similar between languages than other vocabulary.

Synesthetic sound symbolism. Synesthesia as a neurological condition is a strong

coupling of two types of sensory information (Ramachandran and Hubbard 2001).

Synesthetes often see numbers or words as distinctly colored, taste sounds, or see

moving colors when they hear music. The term is also often used in relation to cross-

sensory metaphors, and it is in this sense, rather than the pathological sense, that it is

used here. Synesthetic sound symbolism is the imitation with speech sounds of non-

acoustic phenomena. Generally this means that certain vowels, consonants, or

suprasegmentals consistently represent certain properties such as size, shape,

brightness, texture, speed, etc. The representation is not arbitrary, but is based on some

similarity between the sound and the sensory signal it symbolizes. For example, there

is commonly thought to be an association of palatal consonants and front and/or high

vowels, which have relatively high frequencies (or small wavelengths), with diminutive

qualities, small animals, or objects. This pattern has been documented across a large

number of languages, and although it is not without exception, has long been a

candidate for a universal sound symbolic pattern. Another such cross-cultural example

is the association of rounded vowels with rounded objects. These are the types of sound

symbolism which would explain why subjects in Brown et al. (1955) were able to match

sensory foreign word pairs with their English equivalents. While synesthetic sound

symbolism in general is less likely to be universal than corporeal or imitative sound

symbolism, it is more interesting to investigate. Representation through metaphor is a

very powerful tool, so if humans consistently represent certain properties with certain

sounds this would have helped early humans develop a common vocabulary.

Conventional sound symbolism. This is the association of certain phonemes,

consonant clusters, or syllables with particular meanings. In the literature this

phenomenon is often called clustering, and the units of sound are called phonesthemes.

So far all languages that have been examined have been shown to have phonesthemes

(Bergen 2004), and English is no exception. For example, the words glitter, glisten,

glassy, glow, glimmer, glint, gleam, glance, glare, glower, glimpse all share the cluster gl and

a meaning that has something to do with light or sight. Clash, bash, smash, crash, splash,

lash, gnash, mash all share the syllable ash and their meaning involves violent impact.

Flutter, fly, flit, float, flap, flip, fling all share fl and a meaning that has something to do

with movement through the air. These phonesthemes do not have strict one-on-one

correlations with a meaning however. Glove has nothing to do with light, cash has no

element of impact, and flavor has no connotation of movement through the air. Along

the same lines, conventional sound symbolism can be cross cultural, but in general is a

language-specific phenomenon. At the same time, connections between phonesthemes

and meaning do seem to have psychological reality for speakers within a common

language. Phonesthemes do have a direct connection to meaning, but they are different

from other types of sound symbolism in that, as best we can tell, the brain creates this

connection rather than because of any property intrinsic to the sound.

2.2 Language-Specific Sound Symbolism

The focus of this paper is primarily on synesthetic sound symbolism. This type

of symbolism is extremely interesting because the connection between meaning and

sound is more indirect than purely imitative or corporeal sound symbolism, and yet as

we understand more about how the brain functions, we can begin to understand the

mechanisms behind it. Also, as mentioned previously, finding that synesthetic sound

symbolism has universal elements provides some insight into the evolution of language

whereas we already suspect that imitation of sounds could have had a role in this

process. A review of conventional, language-specific sound symbolism, however, is

also critical to a search for universal sound symbolism because it is conceivable that the

two could share common mechanisms. Therefore this section reviews what we know

about language-specific symbolism and its psychological basis.

Unlike universal sound symbolism, conventional (language-specific) sound

symbolic systems are very well documented. Phonesthemes are well attested not only

in English, but also in a diverse group of languages including a number of African

languages such as Gbaya, Ijo, Hausa, Kisi, Nembe, Igbo, and Yoruba (Childs 1994),

Finnish (Austerlitz 1994), and Japanese (Hamano 1998). Phonesthemes further have a

demonstrable psychological reality which is similar to that of morphemes. Bergen

(2004) showed that English phonesthemes, at least word initially, had priming effects

very similar to morphemes and distinct from the priming effects seen with unrelated

words, words with shared onsets, words with shared meanings, and even words

containing pseudo-phonesthemes. Pseudo-phonesthemes were defined as shared

onsets in words with a related meaning where only a very small number of words

(usually only two) shared that onset and semantic relationship. The fact that

phonesthemes more successfully primed subjects than pseudo-phonesthemes allowed

Bergen to conclude that the frequency of a phonestheme, or the number of words

sharing a related meaning in a group sharing the same onset, is sufficient to give it a

morpheme-like status.

Frequency may also play an important role in the origin of phonesthemes. It is

unclear exactly how phonesthemes come about, but one plausible hypothesis has been

labeled the “snowballing effect” (Blust 2003). Under this model, a group of words

sharing the same phonestheme (hereafter called a phonesthematic group) begins with

two words with related meanings coming to have a common group of phonemes. This

could happen historically by the existence of two words from the same etymological

root, by the creation of a new word through blending, or through borrowing. If the

resulting pair of words has high token frequency, then the shared phonemes may begin

to be psychologically associated with the shared meaning. In addition, it seems that if

the shared phoneme or cluster of phonemes is relatively rare in a given language, then

it is more likely to become psychologically sound symbolic (Austerlitz 1994). The

sound-meaning connection may further start to exert an influence on other words that

share the same phonemes and gradually draw meanings into alignment with the group.

They may even change the form of words with similar meanings and some similar

sound element. Given the importance of frequency demonstrated by Bergen (2004), the

larger the group becomes, the stronger these forces should become. In already

established groups of phonesthemes, both of these effects can be observed. Glory which

primarily means “praise, honor, or admiration accorded by common consent” also at

some point developed an association with the light of heaven. Bergen references an

example of sound change for the group flag, drag, lag, sag which share the meaning of

“slow, tiring, tedious motion.” Sag was originally sacke until the sixteenth century, at

which time it went through an irregular sound change and made this group one word

larger. Exactly how a phonesthematic group becomes psychologically established,

however, is still a mystery. While high token frequency of an originating two words

may be enough to initiate these effects, it could equally be the case that three or four

words of common etymology are necessary to seed a phonesthematic group.

The important thing to notice about phonesthematic groups is that they do seem

to grow by attracting new members. When we combine this with the observation that

phonesthemes seem to be shared between related languages, but not in a wider context

(Blust 2003), this seems to agree more with a model where the brain creates arbitrary

sound-meaning connections than one in which these connections have some iconic or

biological basis. This is not to say that iconic or biologically motivated sound-meaning

correlations do not exist. Some phonesthemes may in fact have a biological basis and a

more universal distribution. Rather the majority of phonesthemes are not universal and

seem to demonstrate that the brain has a bias towards associating sound with meaning.

2.3 Universal Sound Symbolism

The ultimate goal of this paper is to come to some tentative conclusion as to

whether universal sound symbolism can exist and if so what form it could take. Before

wading into the evidence, it will therefore probably be useful to outline what

“universal” might indicate and why it has come to indicate more than one thing. One

very straight-forward understanding of universal sound symbolism is that we must be

able to identify at least one sound symbolic pattern which is present in every language.

This need not be a particular phoneme or cluster correlating with some meaning, but

could potentially be a suprasegmental or a contrast between suprasegmentals which is

realized with different concrete phonemes in different languages. Ideally the pattern

would be applicable uniformly throughout language, that is wherever a word was

semantically related to meaning A, it would incorporate the universal sound pattern B.

This is not a requirement, however, and if the pattern was restricted to a certain class of

words, for example onomatopoeia, it would still qualify as a sound symbolic universal.

However the pattern is realized, in order for it to be universal, there may be no

exceptions, and there certainly cannot be counterexamples. This perspective on what

universal sound symbolism should denote is espoused among others by Gérard

Diffloth, who justly criticizes “the incorrect use of the term ‘universal’ to mean simply

‘found in a number of languages’” (Diffloth 1994: 107). As this view is concerned with a

universal pattern found in human languages, I will refer to it in the future as linguistic

universal sound symbolism.

At this point the reader will be wondering what non-linguistic universal sound

symbolism might be an and why it is relevant. Many studies of sound symbolism have

focused on the psychological element of it, or in other words how subjects assign

meaning to various sounds or sound contrasts. What I will call universal sound

symbolic intuition would involve at least one sound symbolic pattern which is intuitive

to speakers of every language. There might be variation from person to person with

regard to how strongly the pattern is intuitive, and some people tested individually

may not seem to have any intuition about the pattern at all. In any reasonably large

sample of people, however, there should be a significant tendency to map certain

sounds onto certain meanings. This understanding of universal sound symbolism, with

an emphasis on the psychological phenomenon instead of an observed pattern among

natural languages, is typically held by researchers who hypothesize that sound

symbolic intuition has an innate basis (Ohala 1994, Ramachandran and Hubbard 2001).

The degree to which universal sound symbolic intuition, assuming it exists,

would have influenced sound symbolism in natural languages is unknown. There are

multiple possibilities:

1) It may have had a profound influence on the origin of language and may have

constrained language evolution and language acquisition since that time

(Ramachandran and Hubbard 2001). Such a scenario would likely produce

linguistic universal sound symbolism.

2) On the other hand, sound symbolic intuition could be largely directed at non-

linguistic sounds, and could have had no influence on language evolution

whatsoever. This would leave only language-specific sound symbolism, which if

it had any similarity to universal sound symbolic intuition would be due to

chance.

3) A third possibility would be that sound symbolic intuition was initially

influential in the creation and establishment of human language, and may

continue to have a limited influence when language is deliberately created, but

that typically it is directed more towards non-linguistic sounds. In this scenario,

the original language or languages would have displayed universally intuitive

sound symbolic patterns, but the patterns would have gradually changed and

eroded through the process of regular sound change. The patterns are unlikely

to still be recognizable today, but we would observe the sound symbolic patterns

incorporated sporadically here and there in cases where people have put thought

into how a word should sound. This might be most noticeable in naming

systems, diminutive suffixes and other suffixes that denote some characteristic,

and possibly slang.

4) A final unsettling but realistic possibility is that universal sound symbolic

intuition could have a weak or moderate effect on language evolution, but that

patterns in a speaker’s native language could also have an effect on people’s

intuitions (Maurer et al. 2006). Sound symbolic intuitions could still be

considered universal if they were innate, or if they were learned experientially

and were experimentally recoverable in non-linguistic contexts. Patterns in

language could change through regular sound changes, however, so we would

not find truly universal sound symbolic patterns. Furthermore, in languages in

which sound symbolic patterns matching the speakers’ innate intuitions were

absent or reversed, speakers tested for universal intuitions might not

demonstrate them.

In all of the above scenarios universal sound symbolic intuitions exist, but as we can

see, the extent to which they are incorporated into natural languages could vary

considerably. If we are interested in whether universal intuitions have shaped some of

the sound symbolic patterns that we observe in language, however, a minimum

requirement is that the patterns be represented above what would be expected by

chance.

Having outlined the possible interpretations of universal sound symbolism, it is

now possible to review the evidence for and against it. This evidence generally falls

into three lines of research which will be discussed at length below. One involves

magnitude symbolism and an ethological explanation for it called the Frequency Code

Hypothesis. The second is often called either the Maluma-Takete phenomenon or the

Bouba-Kiki phenomenon, but is probably better labeled shape symbolism. The third

line of research has addressed the existence of sound symbolic patterns in general

without focusing on a specific pattern. Recent research into the neurological basis of

synesthesia offers some hypotheses as to the mechanisms behind these types of

synesthetic sound symbolism.

2.3.1 Magnitude Symbolism

The first experimental work concerning magnitude symbolism began in the

1920s. Sapir (1929) presented about 500 English speakers and seven Chinese speakers

with words that contrasted vowel sounds such as mil and mal and asked them to judge

which one was larger and which smaller. He found that when words with [i] were

contrasted against words with [a] at least 80% of his subjects felt that the word

containing [a] was larger. Newman (1933) continued this work, discovering that

articulatory position of the vowel and the acoustic frequency were both important

predictors of the relative size subjects attributed to vowels. In other words, higher

frequency or front/high vowels tended to be judged smaller. He was further able to

show that subjects correlate brightness and darkness with vowels in the same way that

they correlate magnitude.

Given the significance of vowel contrasts for English speakers, Newman also

examined the distribution of vowels among size-related words in English. He compiled

all of the words in Roget’s Thesaurus that were listed under Greatness, Smallness, Size,

and Littleness and divided them into words evoking largeness and smallness. Separate

judges removed words that were ambiguous for size. However, when the word lists

were compared, there was no significant distribution of vowels. Brown (1958) repeated

this procedure, but accounted for usage frequency and still found no sound-meaning

correlation. The implications of these studies seemed to be that magnitude symbolism

was a largely psychological phenomenon, at least among English and possibly Chinese

speakers. It was further curious because was not based on a subconscious knowledge

of vowel distributions in one’s native language.

Despite the lack of a correlation in English, scholars continued to show an

interest in vowel distribution in natural languages. Jesperson conducted an informal

survey in 1933 of a number of languages that highlighted many examples of the

association between [i] and diminutive qualities (Nuckolls 1999). A thorough and

extensive cross-linguistic survey did not appear until the 1970s. Ultan (1978) surveyed

136 languages for size ablauting, or a change of one phoneme or tone that indicates a

difference in size or in a semantically related quality such as distance or number. He

found that 27% of the languages he sampled had size ablauting, while 33% had distance

ablauting. Almost 90% of the size ablauting systems and 85% of the distance ones

conformed to the principle that the diminutive quality was represented by the higher

frequency sound. Other semantically and grammatically related concepts such as

quantity, number, force, intensity, proximity, and weight among others were also

expressed by ablauting but at much lower frequencies.

In a more recent study, Woodworth (1991) studied magnitude symbolism in

deictic pronouns using 26 languages chosen to be maximally genetically distant and

representative of world’s languages. Although not as extensive as Ultan (1978), this

sample of languages was more carefully constructed to include diverse language types.

Woodworth found that in a little over half of the languages, the proximal (i.e. “this”)

and distal (“that”) pronouns had clear vowel contrasts which could be analyzed with

respect to relative frequency. In other cases she was unable to make comparisons

because there were more than two deictic pronouns in some languages, there were

multiple vowels that differed between the two proximal and distal forms, or the

proximal and distal forms had different numbers of vowels. Among those languages

that could be analyzed, however, Woodworth found that proximal pronouns used

higher frequency vowels than distal pronouns in about 70% of the languages. This

pattern was reversed in only about 10% of the languages and the two pronouns had

vowels of equivalent frequency in the other cases. She found very similar results for

place adverbs (i.e. “here” and “there”) and directional affixes (indicating toward or

away from the speaker). Ultan and Woodworth both give compelling evidence that at

least in some areas of language, magnitude symbolism appears to conform to the

pattern described by Sapir and Newman with surprising frequency. The pattern is not

universally represented, but nonetheless is represented beyond what would be

expected from a chance distribution of vowels.

Berlin (1994) extended the study of size symbolism to ethnozoological

nomenclature. He studied bird names in four languages spoken in South and Central

America: Huambisa (a Jivaroan language spoken in north central Peru), Wayampí (a

Tupian language), Apalái (a Cariban language), and Tzeltal (a Mayan language). For

each of these languages he divided the names into those birds that were greater and less

than 10 inches long. He found that small birds tended to receive names containing

higher frequency sounds while larger birds received names with lower frequency

sounds. In Huambisa, for which Berlin had the most information, when he removed

birds with clearly onomatopoeic names, 71% of the birds under 10 inches long had

names containing [i]. Fish names in Huambisa appeared to conform to the same

pattern. Berlin’s findings are interesting because ethnozoological nomenclature is an

area of language where someone or a group of people often consciously give thought to

what a certain animal should be called. The results suggest that when people are

creating new names for things around them, they may make use of sound symbolic

intuition.

Sound symbolism is very appealing in the corporate world because it suggests

the possibility of brand names that will convey certain properties about a product.

Consequently it is no surprise that consumer research has fleshed out some of the

details of magnitude symbolism, both in terms of what qualities fall under the same

sound symbolic pattern as magnitude and in terms of what sound contrasts can evoke

those qualities. A study by Klink (2000) investigated whether sounds that indicate

something is smaller to English speakers could also indicate that it is lighter (relative to

darker), milder, thinner, sharper, softer, faster, colder, more bitter, more feminine,

friendlier, weaker, lighter (relative to heavier), and prettier. He hypothesized that

between two contrasted sounds, the higher frequency sound would indicate these

qualities. Although he did not attempt all possible sound pairings, he found evidence

supporting most of his claims. Words containing front vowels were judged to match

the above qualities better than words with back vowels. Voiceless stops evoked these

qualities better than voiced stops. Fricatives were somewhat, but not always better than

stops and the same was true for voiceless and voiced fricatives. In general, vowels

consistently gave subjects information about Klink’s list of qualities while consonants

appeared to be a less reliable source of information, at least for English speakers. It is

entirely possible that in other languages, contrasts in consonant frequency are more

likely to symbolize magnitude.

The evidence thus far appears to contradict the existence of linguistic universal

sound symbolism. Ultan (1978) found much support for size and distance ablauting,

but it is important to note that 11% (size) and 16% (distance) of the languages he

sampled had ablauting counter to his proposed pattern. This was true of Woodworth’s

study as well. Diffloth (1994) documents such a counterexample in Bahnar (a Mon-

khmer language of Vietnam), where high vowels represent large things and low vowels

small ones. Interestingly, he points out that the system in Bahnar could still be

considered symbolic as size correlates with the degree to which the tongue fills the

mouth rather than with frequency as has been reported in so many other cases. This

may add support for the notion that such systems are formed in agreement with

cultural or universal notions about sounds and their symbolic value.

2.3.2 The Frequency Code Hypothesis

Magnitude symbolism is best explained by the Frequency Code Hypothesis

which has largely been developed by Eugene Morton and John Ohala (Morton 1994,

Ohala 1994). This hypothesis looks to other animals and evolutionary biology in order

to explain why people would associate large magnitude with low frequency sounds and

small magnitude with high frequency sounds. In particular, the hypothesis considers

fundamental frequency, which is the frequency of the sound produced by the vocal

cords before it is modified in the vocal tract. Morton (1994) best describes the

motivation for this pattern in other animals. It probably originated in reptiles and

amphibians, where individuals continue to grow throughout their lifetime and larger

animals are able to win competitions for mates and other resources. Since smaller

animals nearly always lose fights with larger animals and may become seriously injured

during such fights, it is to their advantage to avoid them. They can judge the size of

potential opponents visually, but the calls of larger, more dominant animals have a

lower fundamental frequency than the calls of smaller animals, therefore calling

frequency is also an honest indicator of size. Individuals that correctly identify low

calls with large individuals are likely to survive to an age when they will themselves be

large enough to mate. And indeed, studies have shown that playing the call of a larger

toad will cause smaller toads to retreat (Davies and Halliday 1978), demonstrating that

sound alone is adequate to indicate size.

Among birds and mammals, growth stops at sexual maturity and fundamental

frequency is often, but not always correlated with body size (Ohala 1994, Morton 1994).

However, it still seems to play a role in the communication of size and size-related

motivation. For example, animals displaying aggression usually give harsh, low

frequency calls, while in friendly, appeasing, or fearful situations they tend to give tone-

like, higher frequency calls (Ohala 1994). Large animals will often still be the more

aggressive and dominant ones, so that in these more recent animal lineages it remains

advantageous to correctly associate frequency with size and dominance.

There is further compelling evidence that acoustic frequency has played an

important role in human evolutionary history. This has to do with formant frequencies,

which are another element of vocal quality that indicate body size even when

fundamental frequency does not. Formants are the result of sound filtered by the vocal

tract, allowing only certain frequencies to pass through. They are determined by the

length and shape of the vocal tract involved, which is in turn proportional to body size

(Fitch 2000). In human males, unlike other primates, the larynx descends at puberty,

making the vocal tract longer and the males’ formant frequencies lower. This change is

accompanied by other developments such as facial hair and broadening of the

shoulders which serve to increase the male’s body size or the impression of body size.

Males that could communicate larger size effectively to others should, as among other

animals, have been able to assert dominance and win mates without necessarily having

to fight for them. The descent of the larynx was therefore probably selected for because,

like other indicators of size, males with these characteristics were more successful and

had more children (Ohala 1994). This hypothesis is supported by evidence from birds

that have developed tracheal elongation. The trachea in such birds is often looped or

coiled within the bird’s body, significantly decreasing their formant frequencies. There

is some evidence that these elongated tracheas help these birds sound larger, especially

at night or in heavy vegetation (Fitch 2000). Thus the descended larynx in human males

seems to indicate that frequency is built into human anatomy and evolutionary history.

In order for males with low frequency voices or low formant frequencies to be

selected for, other humans must have been able to associate frequency with size and

dominance. Ohala (1994) offers this as justification for why the frequency code must be

innate in humans, however, it is also possible that it is learned through experience, as

long as it is learned by virtually everyone. This is actually very plausible. Frequency is

generally an honest indicator of size, even among non-living things. For example,

avalanches make a lower sound than small rocks sliding down a hill, and anyone who

has blown through two tubes of different lengths knows that the longer one produces a

lower sound. All humans are likely to have experiences of associations of this type,

whereas the inverse relationship is unlikely to occur often if at all, so it is entirely

plausible that everyone learns this kind of association. There is in fact some evidence

supporting learning over an innate frequency code in that children have been found not

to consistently associate pitch with size until about age 11 (Marks 1987). It seems very

likely that associations between frequency and magnitude would be universal, although

this has not been rigorously documented, but it remains to be seen whether these

associations have an impact on language and intuition in a linguistic context.

2.3.3 Shape Symbolism

Shape symbolism research has centered around matching drawings with sound.

The basic design, developed by Kohler in 1929 and later by Fox in 1935, has been to

present subjects with two line drawings, such as the two pictured below, along with

two words and ask them which word best represents each drawing.

In Kohler (1929) the two words were baluma and takete. Fox (1935) changed baluma to

malumba to avoid obvious similarity with the English word balloon. In a 1947 version of

the experiment, Kohler adopted the final forms maluma and takete. This last example

gives the two words a more consistent composition: continuants and rounded vowels

contrasted with stops and unrounded vowels. In any case, in all three version of the

experiment, speakers of English and German overwhelmingly matched the former

word to the rounded shape and the latter word to the pointed one.

In the quest for universal sound symbolism, this experimental design was

attempted with speakers of various foreign languages. Among these studies, Davis

(1961) repeated the experiment with school children near Lake Tanganyika who spoke

Kitongwe (a Bantu language) and Swahili. Maluma was once again changed, this time

to uloomu because of similarity to a word in Kitongwe. Davis found that the children

did not match the rounded picture with uloomu quite as often as the English school

children that he tested, but that they still had a significant tendency to do so. Another

very informal test with Songe speakers in Papua New Guinea, however, did not find

that subjects matched pictures and words with any level of significance (Rogers and

Ross 1975). Roughly half of the subjects matched maluma with the pointed shape and

half with the curved shape.

According to Westbury (2005), all of these studies on shape symbolism have

shared two main flaws. These were the use of just two words to generalize about

symbolic sounds, and the use of an experimental design transparent to subjects.

Westbury bypassed these problems by using a lexical decision task similar to that used

in Bergen (2004). While it is not clear whether the effect demonstrated by Kohler and

others is dependent on consonants or vowels or both, Westbury chose to focus on a

theoretical association between stops and pointy shapes and continuants and rounded

shapes. Thirty subjects were asked to determine whether a string of letters was a word

or not. The string was centered within a frame that was either spiky or curvy around

the outside, and the experiment measured whether the shape of the frame interfered

with the decision task. Interestingly, there was an interference effect for non-words but

not for real words such that subjects took significantly longer to identify strings of

continuants and vowels in spiky frames than in curvy frames and the reverse was true

for strings of stops and vowels. Westbury’s work not only confirmed that people

generally find stops representative of spikiness and continuants representative of

curviness, but also introduced the interesting idea that words are affected differently

than other sounds by sound symbolic intuition.

One important question about sound symbolism concerns whether it is based on

innate neural connections or connections that are learned early in life. While size-

frequency connections seem to develop in late childhood, this is not necessarily the case

for other types of associations. A very recent paper examined this issue using shape

symbolism to test the judgments of 2.5 year olds (Maurer et al. 2006). The authors found

that toddlers, like the college students in their study, matched rounded shapes with

names that had rounded vowels significantly above chance. Interestingly, toddlers did

not match rounded shapes with rounded vowels quite as often as adults, suggesting

that they may be learning about shape-sound connections and polishing their

associations of this type. Unfortunately toddlers already have significant linguistic

experience, so this study does not show that shape symbolic intuition is innate, but it

does indicate that if it is not innate, it must be learned very early in life. This differs

from previous findings that children do not develop this intuition until about age 9

(Davis 1961), and suggests that when age-appropriate methods are used, very small

children may be able to reliably make such discriminations. This also calls into question

the study suggesting that children do not consistently associate size and frequency until

age 11 (Marks 1987), and suggests that alternative methods might find that younger

children also relate size and frequency.

2.3.4 General Studies

Brown et al. (1955) was actually not the first experiment in which English

speaking subjects were asked to match a pair of foreign, sensory opposites to their

English equivalents. The original idea that subjects should do well at such a task if

sound symbolism were represented in natural languages occurred much earlier and

was repeatedly tested in a series of largely unpublished studies. These included one by

Tsuru in 1934 using Japanese word pairs, followed by Müller (1935) with Swahili and

Bantu word pairs, Allport (1935) with Hungarian pairs, and Rich (1953) with Japanese

and Polish pairs (Brown et al. 1955). All of these studies found that subjects performed

at rates over 50%, although not all of them tested for statistical significance. Brown et

al. (1955) was an attempt to expand the languages that had been tested in this manner

and to correct methodological issues which could have accounted for the above-chance

results. The most important of these corrections was to have someone who was

unfamiliar with the experiment translate the list of English word pairs into the tested

foreign language. This precaution was designed to minimize the chance that the

translator would choose translations on the basis of their similarity to English. If the

foreign words reliably resembled their English translations in just a few cases, this

could potentially be enough to allow the test subjects to perform at levels above chance.

The authors also tested for the effects of presentation by giving one group the test in

written form and having a native speaker read the words, while giving a second group

only the written format.

Brown et al. (1955) was therefore a much more rigorous incarnation of the word-

pair matching experiment than previous versions, but since the overall success rate

continued to be only slightly above chance, around 55-60% depending on the language,

questions of methodology continued to arise. Maltzman et al. (1956) introduced a new

aspect whereby subjects were asked to match an antonymic pair of Croatian words with

their translations in Japanese. They compared this to the normal condition where

subjects matched English word pairs to their equivalents in Japanese and Croatian, and

they found that subjects only performed at above-chance levels when the English word

pairs were involved. Brackbill and Little (1957) further changed the procedure by using

a list of 50 high frequency words rather than pairs of antonyms. These were not strictly

sensory words, as in other experiments, but also function words like when, first, this, etc.

They translated their list into Chinese, Japanese, and Hebrew, and then paired half of

the words in each language with their translation in another language, and half with a

word that meant something different. Subjects were presented with two paired words

and asked whether they meant the same thing or not. They were told that half of the

words were paired with their correct translation. Not surprisingly, subjects failed to

guess correctly at a level above chance.

Brown and Nuttal (1959), in reviewing the three methodologies conclude that

successful guessing must be based on the knowledge of how various phonetic spectra

relate to sensory spectra.

For example, assuming the frequency code hypothesis to

describe reality, if we were considering the pair big-small, we would know that words

meaning small should have higher frequency sounds than words meaning big. If we

were presented with the word pair kit-ket, we would know to pick the word containing

i. However, if the contrast was between ket and kat, then we would know to pick the

word containing e. If we were only given ket and asked whether it meant big, we would

have no way of knowing for sure whether the sounds in the word were of relatively

higher or lower frequency than its antonym. Similarly, if we were given the pair kit-sot

and set-kat, but we were not told that these were both translations of small-big, then we

would not know whether to pair the words such that the first letters matched or such

that the vowel contrast was the same. The authors suggest that because English

speaking subjects do somewhat better than chance when presented with the Brown et al.

(1955) method, but do not when presented with the Maltzman (1956) and Brackbill and

Little (1957) methods, cross-cultural sound symbolic patterns must operate by the

association of a phonetic spectrum, or spectra, with a sensory spectrum. This also

explains how languages with different phonemic inventories could all hold to the same

sound symbolic pattern—as long as they have some of the sounds along the relevant

phonetic spectrum, they could follow the same sound symbolic pattern as another

language using different phonemes on that spectrum.

Brown and Nuttal (1959) discuss phonetic spectra, but the sensory spectra could just as easily be

represented by the presence or absence of a single feature. A spectrum is a good descriptor of magnitude

symbolism and the frequency code, but in shape symbolism the distinction is between stops and

continuants or between rounded and unrounded vowels, and there are no gradations between these

features.

2.3.5 The Neural Basis of Sound Symbolism

Although the synesthetic in synesthetic sound symbolism refers to cross-modal

metaphors rather than clinical synesthesia, studies of the latter have actually begun to

provide some insight into potential mechanisms behind sound symbolism.

Connections have been made between clinical synesthesia and sound symbolism at

least since the 1950s, however, Brown (1958) argued that the two should not be

associated because individual synesthetes of the same variety vary widely in their

associations. For example, if two synesthetes see colors when they hear chords, one

might find that a C major chord is green while the other finds it to be gold. Therefore it

seems unlikely that sound symbolism is a very common, mild form of synesthesia

because it shows a relatively high level of consistency, at least among English speakers.

There is growing evidence, however, that sensory modalities are not the discrete,

separate modalities that they were once thought to be. In a 2001 review, Shimojo and

Shams document a number of cases in which sensory regions in the brain have been

shown to be relatively plastic. For example, in early onset deaf individuals the regions

of the brain normally devoted to hearing are often used for vision in addition to the

normal visual areas. Furthermore, the primary and secondary visual cortical areas are

activated by reading Braille in blind but not sighted subjects. Sensory modalities have

also been shown to interfere with and alter other sensory modalities. This is commonly

experienced with a visual cue altering the perception of location or quality of a sound,

as when a ventriloquist makes it appear that someone else is speaking. However,

sounds can also alter visual input. For example, if multiple beeps accompany a single,

brief flash of light, subjects will perceive the flash as multiple flashes. Sounds are also

used to disambiguate visual stimuli. If in a cartoon, two balls travel towards each other,

fuse, and then separate, subjects will see the two balls bounce off each other if there is

an accompanying sound of appropriate sharpness and duration. If there is no sound

they will see the balls travel past each other. All of this is just to say that the senses are

relatively interconnected in normal people so that hypotheses about sound symbolism

that involve cross-modal neural connections may not be that far fetched. The cross-

modal neural connections underlying sound symbolism would not have to be the same

as those that we think are unnaturally strengthened in synesthetes, but synesthesia has

nonetheless inspired hypotheses about the neural basis of sound symbolism because it

is a well established example of cross-modal connection.

The major hypothesis regarding the neural basis of sound symbolism was

presented as part of a larger paper on synesthesia by Ramachandran and Hubbard

(2001). They propose that there are a number of non-arbitrary connections in the brain

between the maps of speech lip and tongue movements and the representations of

certain phonemes or visual qualities in auditory and visual areas. They present a

number of observations to support the existence of such sensory-to-motor connections.

The most familiar observation is probably the translation of audio input into movement,

or dance. Another observation is that there are rare forms of synesthesia in which

particular sounds evoke the automatic adoption of specific postures. If we suppose that

synesthesia is an unnaturally strong form of normal neural interconnectivity, then this

would support the existence of sensory-to-motor connections. Finally, Ramachandran

and Hubbard cite the existence of mirror neurons, which fire in humans and monkeys

both when they are watching someone perform a task with their hand or mouth and

when they themselves are performing it. Mirror neurons have also been shown to fire

when someone is listening to sounds that are easily identified with the task, including

speech sounds (Rizzolatti and Craighero 2005). Given the existence of neurons that fire

in audio, visual, and motor contexts, the existence of innate, non-arbitrary connections

between audio and visual input and the motor map for the production of corresponding

speech sounds seems possible. As discussed earlier, however, there is no reason that

these connections would have to be innate. It is equally possible that they could form as

a result of experiential learning during childhood, as long as the sound-quality pairing

was found in the natural world.

2.3.5 Conclusions from Sound Symbolism Research

The literature suggests that linguistic universal sound symbolism, that is a sound

symbolic pattern expressed in natural languages without exception, probably does not

exist. Magnitude symbolism, described by the frequency code hypothesis, has very

compelling reasons to be universal, but we have yet to find an area of language in

which there are not exceptions and counterexamples to it. Vowel frequency, the

strongest indicator of size to English speakers, does not seem distributed such that

words affiliated with smallness have higher vowel frequencies. Magnitude is widely

expressed in size ablauting systems, and deictic pronouns, but there are exceptions and

counterexamples. The frequency of vowels in bird names in some languages

corresponds to the birds’ size, but this is not true for all bird names. In every way that

we find magnitude symbolism incorporated into language, it always has exceptions,

and this suggests that if we find it incorporated in other aspects of language, there will

be exceptions there as well. Shape symbolism in natural language has not really been

studied, but if speakers of Songe really do not have intuitions matching those of

English, German, and Kitongwe speakers, then that suggests that their language will

not follow the same pattern either. We cannot rule out linguistic universal sound

symbolism, but if all hypothetically universal sound symbolic intuitions are

incorporated into language by a similar mechanism, then the evidence suggests that

that mechanism will always lead to some exceptions.

Universal sound symbolic intuitions seem likely, but are far from proven. The

frequency code hypothesis is well supported by evidence from other animals and

human anatomy. It seems likely that even if intuitions about size and frequency

weren’t found in other animals and weren’t innate, that humans would learn about

them from their environment. The evidence that magnitude symbolism is expressed

widely across languages also supports the notion that humans have matching

widespread intuitions. It would be hard explain the prevalence of magnitude

symbolism in language if it were not motivated by some psychological factor.

Currently there have been studies on speakers of relatively few languages, however,

and a much wider cross-linguistic sample would be needed to corroborate the

frequency code hypothesis. Shape symbolism is similarly in need of a wider cross-

linguistic sample. The evidence that speakers of Songe do not share intuitions about

shape symbolism with English speakers suggests that this pattern may not be universal,

however it does not completely rule out the possibility. Tsur (2006) argues that

normally when we process language, we do not attend to sound quality, but only to the

abstract sequence of phonemes. He claims there is a “poetic” mode of listening which

we can use when we want to attend to sound quality as well as the sequence of abstract

phonemes. To make use of sound symbolic intuition, it makes sense that one would use

the “poetic” mode of listening and consider actual sound quality. In many studies of

sound symbolism, subjects are asked to consider sound quality when they make

judgments about meaning. In the Rogers and Ross (1975) study on Songe speakers,

subjects were only asked to say which drawing they thought was a maluma and which a

takete, not which name sounded like it matched one of the drawings. Subjects may not

have attended to sound quality at all, and given the results they seem to have

responded at random. A more careful experiment with speakers of this language seems

warranted, and along with a wide survey of speakers of different languages could give

us a better idea for whether intuition about shape symbolism is universal.

One issue that remains unclear is the extent to which sound symbolic intuitions

have been incorporated into natural language. If they have acted as a general constraint

on language evolution, then we should expect to see sound symbolic patterns

throughout language. The patterns should not only show up in the somewhat

specialized contexts like naming or ablauting systems where at least magnitude

symbolism has been largely found so far. They should be apparent in words

throughout the lexicon, including the adjectives that label the sensory qualities being

symbolized (i.e. big and small in the case of magnitude symbolism). At the moment this

does not seem to be the case. Two attempts uncover a constraint on size-related words

have found no significant distribution of high frequency vowels. But before concluding

that sound symbolic intuition does not act on sensory labels, it will be worthwhile to

reconsider the evidence from Brown et al. (1955).

3 Sound Symbolism and Guessing Foreign Word Pair Meaning

Brown et al. (1955) is a fascinating study because it suggests that speakers’ inner

sound-meaning associations are reflected by patterns in natural languages. It is not

difficult to understand why English speakers would all have similar sound-meaning

associations such that they would match foreign word pairs with their translations

consistently. Their experiences with language are relatively similar, and they should

have the same knowledge of English phonesthemes or other patterns in the distribution

of sounds in English. It is even understandable how in some cases, people that speak

different languages might have similar judgments about what sounds best imitate

natural sounds, visual and tactile qualities, or size. Although this has not been

satisfactorily demonstrated, there is compelling logic behind the Frequency Code

hypothesis and hypotheses about neural interconnectivity. But the idea that these

sound-meaning correlations are reflected in natural languages to such an extent that

people can correctly translating a pair of antonyms, even when limited to certain

sensory antonyms, is incredible. If there are clear constraints on the sounds making up

sensory adjectives, why don’t these words clearly resemble each other from language to

language? If there are not constraints, how can test subjects assign them meanings

correctly? If subjects in the Brown et al. study were predominantly using honest sound

symbolic cues to match pairs with their translations then this has implications for how

some parts of language are created. Therefore this next section aims to experimentally

determine whether sound symbolic cues are generally responsible for correct

translation of antonymic foreign word pairs.

3.1 Methods

It is most likely that subjects use a combination of different cues when they

match foreign word pairs with their meanings. These could include

a. characteristics of the written word such as length or shape

b. resemblance to an English word

c. sound symbolic patterns specific to English

d. cross-cultural or universal sound symbolic patterns

It is unlikely that subjects use only sound symbolic patterns because in any case that is

not perfectly controlled, they probably use a combination of all of the above cues.

Instead of trying to completely rule (a) and (b), it is better to ask whether sound

symbolic patterns are one of the cues that subjects use to correctly match foreign words,

and possibly if they are necessary cues for subjects to do so. This can be done by

comparing the success on a group of foreign words that display a known sound

symbolic pattern to the success on a group that does not display this pattern. If success

with the “symbolic” group of words is much higher than with the “non-symbolic”

group, this would indicate that sound symbolic patterns contribute to successful

matching. Provided that other cues equally promote correct matching in the two

groups, which they should if they are random, if subjects guess correctly at a frequency

significantly above chance for the symbolic but not for the non-symbolic group, this

would further suggest that sound symbolic patterns are responsible for their success.

The best described sound symbolic patterns are those associated with magnitude

symbolism. According to the Frequency Code hypothesis, smallness and related

qualities are associated with any sounds that are relatively high in frequency (Ohala

1994). In order to simplify the process of assigning words to symbolic and non-

symbolic categories, however, I chose to limit this criterion to vowel contrasts for this

experiment. This is reasonable given that most studies have focused on vowel but not

consonant contrasts (Nuckolls 1999).

The symbolic and non-symbolic word lists were created using translations for

big, huge, small, and tiny in sixteen different languages. Languages were chosen based

on the selection of dictionaries available in the Swarthmore College library. Languages

were only used if they were non-Indo-European to rule out the possibility of cognates

with English and to minimize the chances that experimental subjects might have

studied them. In addition, only dictionaries using roman characters and with

descriptions of pronunciation were used. All one and two syllable definitions for the

four words were copied down in the order in which they were listed in the dictionary.

Three or more syllable words were not included to minimize unnecessary complication.

With large numbers of syllables it becomes unclear what sounds will most effect a

subject’s judgment. Word pairs were created using the first translation for big and the

first translation of small or tiny with the same number of syllables. In this way, words

were roughly balanced for length, although one word might still take up more space on

the page than its pair. This method also prevented me from forming pairs on the basis

of whether I thought they sounded symbolic or not. Instead pairs were based on the

order of dictionary listings and were separated into the symbolic or non-symbolic

group by a consistent metric. Ultan (1978) found that small- big contrasts were not only

denoted by the high, front vowel [i] and the low, back vowel [a], but more generally by

a front vs. back vowel contrast or occasionally by a high vs. low vowel contrast. This

can be illustrated by the English clink/clank, where clank is the heavier sound, but the

vowel is a low, front vowel, not a low, back vowel. In order to capture both such

contrasts then, words were given a score equal to their number of front vowels and high

vowels. High, front vowels were counted once for each feature. If the word meaning

“small” had a higher score than the word meaning “big”, a pair was counted as

symbolic. If the scores were the same, or “big” had a larger score, then the pair was

considered non-symbolic. So for example, in Malay big = besar = 1 point because it has a

front vowel and a low, back vowel. Small = kechil = 3 points because it has a front vowel

and a high, front vowel, so this word pair would be considered symbolic. On the other

hand, in Arabic big = kabi:r = 2 points because there is a high, front vowel. Long vowels

were considered the same as short vowels for this analysis and were presented to

subjects as short vowels. The Arabic small = saghi:r = 2 points again because of the high,

front vowel, so these two words are equivalent and the pair would be considered non-

symbolic. A total of 20 word pairs were compiled and the first eight pairs in each of the

symbolic and non-symbolic groups were used in the final two lists. Of the eight pairs in

each list, seven were two syllable words and one pair contained one syllable words.

In addition to the sixteen foreign word pairs, twenty-five nonsense word pairs

were created and included in the survey. These words were added in the hopes of

diluting any obvious patterns and also to pursue some interesting tangents. Five of

these were designed to test how similarity to an English word would affect how

subjects assigned meaning. Two synonyms of small and three synonyms of big were

each modified slightly. The synonyms of small were contrasted against a nonsense

word with higher frequency phonemes while the synonyms of big were contrasted

against a nonsense word of lower frequency. If subjects chose consistently with the

Frequency Code, they would choose a different word than if they were affected by the

similarity to an English word. An additional ten nonsense words were designed to test

the relative strengths of consonants and vowels as signals of size. English speakers

judge voiceless stops as smaller than voiced stops, and fricatives as smaller than stops

(Klink 2000). Therefore for five words, a word with a voiceless stop and a low or back

vowel was paired with a word containing a voiced stop and the vowel [i]. This was

called the “voiceless vs. voiced stop” group. In the other five words, a word with a

fricative and a low or back vowel was paired with a word containing a stop and the

vowel [i]. This was labeled the “fricative vs. stop” group. Finally, five words were

designed as a positive control to be symbolic with both consonant and vowel contrasts

following the frequency code. The remaining five words were designed as a negative

control to be non-symbolic using vowels and consonants of similar frequencies.

The orthography of all words presented to subjects was adjusted so that native

English speakers would read words with at least approximately correct pronunciation.

For example, the high, back vowel [u] was represented as oo in positions where it would

not be pronounced that way. The Basque orthography tx was converted to its English

equivalent ch. The Korean vowel eu (/∧/) was represented as u, and the rounded front,

high vowel [ü] was written as i. Although in most words [a] would tend to be

pronounced as [æ] by English speakers, there were no cases in which an [æ] instead of

an [a] would greatly change the frequency contrasts in the pair of antonyms, so no

changes were made to this orthography.

The final survey included 41 word pairs randomized for order and for whether

“big” or “small” was presented first. The sequence order was determined from random

numbers generated at http://www.random.org, a web site that generates random

numbers using atmospheric data. The order of the words in each pair was determined

by tossing a coin. The survey was sent by email as a Word document to 60 college

students between the ages of 18 and 22 who were all native speakers of American

English. Subjects were asked to mark the word in each pair that they thought sounded

like it meant “small”. Subjects were asked to report any languages that they knew, had

studied, or were familiar with. The format of the survey can be viewed in Appendix I.

Performance on the foreign word pairs was measured in two ways. A binomial

test was used to evaluate whether subjects matched pairs correctly at frequencies

greater than chance for all foreign word pairs and for the symbolic and non-symbolic

groups separately. To directly compare performance on symbolic and non-symbolic

word pairs, the number of words subjects correctly matched for each group was

compared using a Wilcoxon signed-rank test for non-normal data. This test takes into

account individual variation, in this case by analyzing each individual’s performance on

the symbolic word pairs relative to their performance on the non-symbolic word pairs.

Thus if most subjects scored better on the symbolic word pairs than they did on the

non-symbolic word pairs, this would lead to a significant result even if the mean scores

on the two groups of word pairs were very similar. The five groups of invented words

were not analyzed with any statistical test because of the small number of words tested

in each one. Instead, the average number of “correct” guesses was calculated for each

group and then compared with the other groups. As in Brown et al. (1955), the

percentage of subjects that correctly matched each pair was also analyzed using a

binomial test. Given the large number of statistical tests conducted, I chose to use a

Bonferroni correction such that P-values less than 0.001 were considered significant.

3.2 Results

In total, 56 students completed

the survey. Two subjects were familiar

with Hungarian, and one was studying

Arabic, therefore their answers on

these questions were discarded from

the analysis. Overall, subjects matched

foreign word pairs correctly with their

meaning more often than chance (Table

1). This was also true for the group of symbolic foreign word pairs, but not for non-

symbolic pairs.

The percentage of subjects that matched a given symbolic foreign word pair

correctly ranged from 41.1% to 92.9% (Table 2). This range for non-symbolic foreign

word pairs was 30.4% to 69.6%. In general subjects matched significantly more

symbolic foreign word pairs correctly (median = 6) than non-symbolic word pairs

(median = 4; t = 5.81, df = 55, p < 0.0001).

Table 1. The average percentage of correctly

matched pairs for all foreign words and for

symbolic and non-symbolic word pairs separately

(n = 56 subjects).

Group

Correct

Significance

of Departure

from Chance

Foreign word

pairs

59.9%

p < 0.001

Symbolic word

pairs

70.1%

p < 0.001

Non-symbolic

word pairs

49.8%

p = 0.04*

*Not significant after the Bonferroni correction

Table 2. The percentage of subjects (n= 56) that correctly chose the word

meaning “small.” “Big” is presented first in each word pair.

Word Pair

Language

% correct

“Symbolic”

nagy- kicsi

Hungarian

75.5% *

handi- chiki

Basque

92.9% *

chempo – chooncoon

Tibetan

41.1%

besar- kechil

Malay

64.3%

mwuk- ngin

Trukese

71.4%*

nui- iki

Hawai’ian

76.8%*

babba- tsigil

Hausa

62.5%

nene- nini

Kikuyu

78.6%*

“Non-symbolic”

kabir- saghir

Arabic

58.2%

iri- kicik

Turkish

30.4%*

kuda- jagun

Korean

39.3%

kulu- nono

Bemba

69.6%*

agi- nta

Igbo

39.3%

khulu- ncane

Zulu

62.5%

ge- ah

Cayuga

35.7%

waru- uta

Tarahumara

64.3%

*P< 0.001

Among the invented words, the percentage of subjects “correctly” matching a

given word pair with its meanings ranged from 30.4% to 87.5% (Table 3). The average

percentage of subjects choosing the higher frequency word among symbolic control

pairs was 80.6%. The average for word pairs in the fricative vs. stop group was 76.4%,

63.4% for the voiceless vs. voiced stop group, and 53.6% for the English-like word pairs.

The average percentage of subjects choosing the “correct” word of the non-symbolic

word pairs was 45.0%.

Table 3. The percentage of subjects (n= 56) that chose the word containing the

higher frequency vowel(s) for 25 invented words. For non-symbolic control

words, one word was designated “correct” for scoring purposes. The word

containing the higher frequency vowel, or the “correct” word is presented

second.

Word Pair

% choosing higher frequency vowel

Positive Control: symbolic

balaz- kilsu

73.2% *

godan- ipich

83.9% *

gujre- shekri

76.4% *

orveb- filka

87.5% *

vadong- inkush

82.1% *

Negative Control: non-symbolic

felpo- athek

35.7%

zench- chedj

46.4%

watla- kwona

53.6%

brelko- egralp

30.4% *

ob- da

58.9%

English-like

smar- shil

62.5%

ahob- hugi

58.9%

dabo- ibig

53.6%

dondag- gigank

30.4%*

minag- ziji

62.5%

Voiceless vs. Voiced Stops

pako- bidu

55.4%

fok- thig

67.9%

atin- udin

60.0%

opor- dimir

85.7% *

tras- grish

48.2%

Fricative vs. Stop

yaze- yeggi

73.2% *

cho- twi

75.0% *

shozat- kizta

66.1%

haju- habi

81.8% *

aflet- pliki

85.7% *

*P < 0.001

3.3 Discussion

Subjects in this experiment were able to correctly choose which word meant

“small” at frequencies greater than chance when a word pair was symbolic.

Furthermore, when the pair was non-symbolic, subjects failed to do better than chance.

This indicates that speakers used vowel frequency to make at least some choices. There

are also indications that subjects may have used other types of sound symbolic cues to

correctly translate pairs. Although exactly half of the word pairs had symbolic vowel

contrasts, the overall pairs were matched correctly 60% of the time. Also for 11 out the

16 pairs more than 50% of the subjects made correct translations. Some pairs that were

labeled “non-symbolic” may still have had symbolic consonant contrasts which helped

subjects translate them correctly. For example, kabir-saghir contrasts a fricative with a

stop in the initial syllable which would agree with the frequency code. Although the

percentage of subjects correctly translating this pair was not significant, it did approach

60%. The prominence of contrasts in accented syllables may also have outweighed the

overall vowel contrasts in certain pairs, further contributing to high scores on some

non-symbolic pairs. These factors along with some chance ones may have contributed

to the overall success rate being above 50%.

In addition the results suggest that other criteria which subjects may use on such

tests are not consistent enough among a large group of words to be sufficient for

success. In the case of individual word pairs, there may have been factors unrelated to

sound symbolism which lead subjects to guess correctly. For instance, in the pair waru-

uta both words are very similar aside from the fact that uta is shorter. Word length in

this case may explain the 64% of subjects who guessed correctly on this pair. This is

also true for iri-kicik, where 70% of subjects guessed incorrectly that the shorter word iri

meant small. However, in the context of the larger group of words, word length does

not play a factor because these two word pairs cancel each other out. In fact, there was

no factor or combination of factors which allowed subjects to consistently guess

correctly and do better than chance across the non-symbolic word pairs.

The results from responses to the invented words indicate that conflicting signals

reduce a group of test subjects’ ability to correctly match pairs to differing degrees.

When a fricative-stop contrast ran counter to the relative vowel frequencies in a pair of

words, subjects seemed to ignore this information and focus on the vowel contrasts. A

contrast between voiced and voiceless stops, however, seemed to have a stronger effect,

consistent with Klink (2000). English-like words had the strongest effect in countering a

signal from symbolic vowels. Sharing one syllable with an English word appeared to

be sufficient to keep more than 60% of the subjects from choosing the word with higher

frequency vowels to mean “small.” Sharing two syllables, as in the dondag-gigank pair

appeared to be enough to have a significant percentage of subjects choose against the

frequency code. In the one case where I accidentally created a word containing an

English diminutive prefix, dim-, that also reinforced the information from the vowel

contrasts, the result was a very high score (opor-dimir 86%). These results confirm the

fact that a likeness to an English word could certainly explain why many subjects

correctly translate a certain word pair, and suggest that highly significant results should

be carefully checked for evidence of this caveat.

4 Analysis of Brown Data

4.1 Methods

Part a

Although on average subjects in the foreign word pair test in the previous

section seem to have used sound symbolic cues to match pairs, this does not necessarily

explain the results of Brown et al. (1955). The Brown et al. results were therefore

analyzed to rule out several trivial cues that subjects might have used and strengthen

the conclusion that the results were largely due to concordance with subject’s intuitions

about sound-meaning correlations. Given that resemblance to English can have a

strong effect on judgment, foreign words were examined for resemblance to English

and whether they shared their first letter or an internal consonant cluster with a

member of the English word pair or an obvious synonym. In addition, the effect of

length was measured in two ways. In the first, the data were examine for instances in

which subjects matched the longer foreign word to the longer English word and a

significant percentage of subjects answered correctly. In the second, word length was

examined for word pairs that related to magnitude in any way. While the use of longer

words to represent larger magnitude is a form of symbolism, it is different from the

types of sound symbolism that have been examined so far. Therefore it was counted

among the random factors that might lead a significant number of subjects to translate a

word in the same way, even though it may not be a random factor.

Part b

The word pairs presented in Brown et al. were also analyzed to determine

whether pairs for which subjects did especially well on multiple languages had any

common element. This information would indicate the semantic domains in which we

would be most likely to observe universal sound symbolism. “Especially well” was

arbitrarily defined as 80% of subjects correctly guessing a given pair. In addition, a pair

of antonyms was only selected if subjects did especially well for at least two of the three

languages tested. An average score (% subjects that guessed correctly) across the three

languages tested was calculated for each word pair that qualified.

4.2 Results

Part a

Although there were some resemblances to English words among the dataset,

these explained less than half of the correct choices. In a total of 15 cases, one of the

foreign words presented shared a first letter with its English translation. Brown et al.

(1955) administered the test under two conditions (written and oral presentation vs. just

written presentation), so in total there were 30 instances out of 126 total possible

instances in which subjects were presented with a foreign word that shared a first letter

with its correct translation (see Appendix II). Of these 30 instances, a significant

percentage of subjects gave a correct translation 13 times. In other words, subjects

matched a foreign word with an English translation because they shared a first letter in

no more than 43% of cases in which such a strategy was possible. Of six instances in

which an internal part of the foreign word was similar to its English translation, a

significant percentage of subjects guessed correctly in three of those six cases. Word

length by either measure did not seem to play a large role. In total there were 22

instances in which subjects were presented with a foreign word visually longer than its

antonym whose correct translation in English was also visibly longer than its antonym.

Subjects only matched the longer foreign word with the longer English word in 7 of the

22 instances. Eight of the 21 word pairs used in the experiment were clearly a type of

magnitude. For 8 of the 24 translations of these pairs, the longer word corresponded to

the larger magnitude. However, subjects guessed correctly significantly more than

chance in only four of the cases where length and meaning were correlated. Overall, of

all 44 instances where a significant percentage of subjects guessed word meaning

correctly, 18, or 41%, could potentially be explained by word length or chance

resemblance to English.

Part b

For 5 out of 21 word pairs 80% or more of the subjects correctly translated the

word pair into at least two languages. These word pairs and the average % of subjects

that correctly matched them to English across Chinese, Czech, and Hindi were: hard-soft

(80%), blunt-sharp (77%), bright-dark (73%), fast-slow (70%), and long-short (69%). In

addition, warm-cool (69%) showed significant results for all three languages, but fewer

than 80% of subjects guessed correctly for the three languages.

4.3 Discussion

Part a

The results suggest that the subjects in Brown et al. (1955) successfully assigned

translations to at least some foreign word pairs on the basis of sound symbolism. It is

striking that roughly 40% of their correct responses could be explained by seemingly

random factors, and it is possible that this is an underestimate. There is always the

possibility of a similarity to English which I did not recognize, but which was clear to

those in the study. It is unlikely, however, that all of the correct responses could be

explained by similarity to some English word not apparent to me, especially when the

similarity must have been apparent to a very high percentage of subjects in the study.

In addition, if universal sound symbolic patterns exist, then we would expect there to

be similarities to English among the foreign word pairs. Some of what I am

conservatively assuming to be random similarity might in fact be symbolic. It seems

safe to agree with the conclusion in Brown et al. (1955) that at least in some cases,

subjects must be using sound symbolic patterns in the foreign word pairs to assign

correct translations.

Part b

Interestingly, the five most successful word pairs from Brown et al. were all very

basic sensory terms. This is in contrast to other pairs in the experiment such as

beautiful-ugly, happy-sad, or sweet-sour. Bright-dark, fast-slow, and long-short could all be

considered types of magnitude, so it is not surprising that subjects have strong

intuitions about them. Hard-soft and blunt-sharp are also basic sensory terms that might

be represented by synesthetic symbolism, both of a tactile-auditory sort. Blunt-sharp is

particularly reminiscent of the maluma-takete phenomenon as these adjectives would

describe the pointy and curvy pictures if they were instead three-dimensional. We

might expect that tactile-auditory symbolism would be widespread since the tongue

movements that produce speech sounds would be associated with tactile sensations.

Given that this notion is supported by the high success rates subjects had on the blunt-

sharp and hard-soft pairs cross-linguistically, it may be worthwhile in the future to

examine the distribution of tactile-auditory symbolism in a larger set of natural

languages.

5 Antonym Pair Conformity to the Frequency Code

The results so far are consistent with the idea that the Brown et al. results were

likely due to universal sound symbolic patterns incorporated in natural language.

Subjects did well on the sound symbolic word pairs in section 3, and they were not able

to do well without vowel frequency cues indicating the word that meant small,

suggesting that they have also needed sound symbolic cues to do well in previous

experiments. Whether such cues would be sufficient for subjects to do well on a wider

sample of languages is another matter. This section examines the incorporation of the

frequency code into natural languages since this pattern is the best studied and is the

best defined.

5.1 Methods

The only way to definitively determine how widely the frequency code is

expressed in natural languages is to sample a very large number of them. Even if one

were just randomly choosing one word that meant “small” and one that meant “big”

from each language, this would be a daunting task. In reality, one would need to

sample all of the words meaning “small” and “big” from each language, score them,

account for word length, and somehow give each language a score based on this

information. This would only give us information about adjectives like big, huge,

enormous, small, little, tiny however, and we would not have any idea about whether

other areas of the language have been affected by the putative sound symbolic

constraint. Clearly other ways to estimate the frequency code’s pervasiveness in natural

languages are necessary.

The word list generated for the foreign word pair experiment above can be used

to give a very rough estimation of how significantly the frequency code is represented

among words meaning “big” and “small.” Nothing can be concluded from the list itself

because we have no sense of how often in a randomly chosen pair of words one of the

words will have more front and/or high vowels than the other. Furthermore, the

sample of languages is small, and only one word pair is taken from each language.

Instead, I compared the list with two other lists of word pairs, compiled using the exact

same method and the same dictionaries, with the addition that three and four syllable

words were also used if one or two syllable translated pairs could not be created. For

both of these comparison lists, I chose simple words that I expected to find in all

dictionaries. One list contrasted translations of river against translations of sun while

the other contrasted eat against walk. There is no reason to expect that in either case one

of the words will consistently have a higher score than the other, as measured in section

3.1, so this should provide a baseline from which to see if “big” and “small” deviate

from chance. A χ-squared analysis was used to determine whether the occurrence of

symbolic word pairs in the big-small list was significantly different than in the eat-walk

or river-sun lists. Because the word corresponding to “small” in the eat-walk and river-

sun lists was arbitrary, I determined symbolic and non-symbolic word pairs first with

one word corresponding to “small”, then the other, and used the condition in which the

most word pairs were symbolic for the analysis. This was the most conservative

approach since it decreased the difference in the number of symbolic word pairs

between the three lists.

5.2 Results

Word pairs were translated into a

total of 22 languages, but it was not possible

to analyze the translated pairs in every one

of these languages. It was unfortunately not

always possible to create a translated pair

where each word had the same number of

syllables, and it was unclear how pairs with

different numbers of syllables should be

compared. It was especially difficult to

create translated pairs with equal numbers

of syllables for the nouns as these entries

generally only had one translation. Therefore the total number of translated pairs that

could be analyzed in the big-small list was 20, in the eat-walk list 20, and in the river-

sun list it was 15. The big-small list of word pairs contained a larger proportion of

symbolic pairs than either of the other two lists, but this difference was not statistically

significant (χ

= 2.79, df = 2, p > 0.05; Fig. 1).

In this χ-squared analysis, the null hypothesis is that the distribution of symbolic pairs across the three

lists is due to chance alone. We can calculate an expected number of symbolic pairs for each list, and then

if the observed number of symbolic pairs diverges enough from the expected number, we can reject the

null hypothesis. The expected number of symbolic pairs is based on the total number of languages each

list was translated into and on the average proportion of symbolic pairs to total word pairs. So if in the

big-small list, about half of the pairs were symbolic and in the other two lists about a third were symbolic,

then the expected number of symbolic pairs for each list will be between a half and a third of the total

Figure 1. The number of big-small word pairs

following the sound symbolic pattern described in

section 4.1 was not significantly greater than

chance. The expected number of symbolic pairs is

based on the total number of languages each word

pair was translated into. In total, big-small was

translated into 20 languages, eat-walk into 20, and

river-sun into 15 languages.

Big-Small Eat-Walk River-Sun

Observed

Expected

Number of Symbolic Word Pairs

Word Pair

5.2.3 Discussion

The above result suggests that sound symbolic intuition does not place a

constraint on the creation of sensory adjectives, or that if it does, the constraint is very

subtle. The proportion of symbolic pairs in each list was greater in the big-small list than

in the other two—about half of the big-small pairs were symbolic while only about a

third of the pairs were symbolic in the other two lists. If the proportion of symbolic

pairs in each of the big-small, eat-walk, and river-sun lists remained constant at one half

and one third as the number of languages sampled increased, the difference between

the lists would become significant after analyzing about 50 languages. Even if half of

the big-small pairs being symbolic was significantly greater than what we would expect

by chance, though, this would indicate a very subtle constraint on language evolution.

One artifact of the methods that could have weakened the results was the way in

which symbolic pairs were defined. To begin with, only vowels were considered, so

that if there was a contrast between higher and lower frequency consonants but no

vowel contrast, the pair would have been considered non-symbolic. Somehow taking

consonants into account as well might have increased the number of symbolic pairs in

the big-small list, although it likely would have increased the numbers in the other two

lists as well.

Another factor that should be considered is the effect the sampling method used

to create pairs for the analysis could have on the results. It may be that in many cases

the big-small pairs analyzed were not really a pair in the minds of speakers of each

given language. For example, “big” meaning “famous” may have been listed as a

translation in one of the dictionaries and paired with a word meaning “small.” No one

number of translated pairs in that list. The expected number will be the same for big-small and eat-walk

because these lists were translated into the same number of languages, but the expected number for river-

sun will be smaller because this list could only be translated into 15 languages.

would normally pair these two words together as opposites, and so there might not be a

sound symbolic constraint on them. Similarly, “big” meaning “expansive” could have

been paired with “small” meaning “short in stature.” These are still adjectives that

convey magnitude and might normally follow a sound symbolic pattern when

contrasted with their natural opposites “narrow” and “tall in stature”, but do not

always when they are paired with each other. In fact, if big-small word pairs had been

symbolic significantly more than expected, this would have indicated a strong

constraint on size word creation since it would mean that almost all “big” related words

used a certain subset of vowels and all “small” related words used another subset of

vowels. However, this was not the case, so while we might find a constraint with a

larger sample size, at best it would be a weak one.

6 Sound symbolism, Arbitrariness, and the Evolution of Language

As the above results suggest that sound symbolism is at best inconsistently

incorporated into natural languages, it is worth considering why we would expect it to

be incorporated at all. As mentioned earlier, the idea of universal sound symbolism has

been particularly interesting in the context of the origin of language. In order for

communication between two individuals to work, both need to share the same

understanding of what the signals they are using mean. In moving from a situation

where the is no communication to one where there is communication, participants must

first develop a common set of basic signals. Computer simulations suggest that even if

humans created arbitrary signals to refer to things around them, they could have come

to a common understanding about what those signals meant (Steels and Kaplan 2002)

and subsequently developed a language (Nowak and Komarova 2001). The

development of language in this type of simulation is dependent on a chance event

whereby the signs representing some feature of language come to have a common

element and humans are able to construct mental rules based on the shared common

element. If humans shared common assumptions about how certain sounds should be

used to indicate distance from the speaker or to name certain properties or objects this

would have speeded up the process of coming to a consensus and in some

circumstances might have increased the incidence of chance resemblances. Once

language was established, sound symbolic words would also be processed faster and

would be more easily comprehensible (Bergen 2004, Hinton et al. 1994). Therefore

groups of people that utilized sound symbolism would probably have developed

language earlier and would have used it more effectively once developed, leading them

to become more successful than other groups. The selective advantage of incorporating

sound symbolism into language, especially during the initial process of development,

makes it seem likely that we would observe sound symbolism in language.

On the other hand, there are reasons why arbitrariness in language should be

selected for which would tend to counter the selective advantage of sound symbolism.

One is that not all concepts can be metaphorically related to a set of sounds. We can

imitate simple qualities like magnitude and texture with sound, and perhaps we could

even derive names for things based on their salient qualities, but how should sixty-four,

interesting, or justice be represented? At a certain point, sounds must be arbitrarily

assigned to meanings in order for people to express their ideas. So the ability to create

and learn words composed from arbitrary sounds would be selected for, and in fact

must have been selected for in order for us to have developed a language that can

express complex concepts.

A second selective advantage of arbitrariness, somewhat paradoxically, also

arises from the need for the signs in communication to be shared between participants.

In order for any type of communication to function, all individuals involved must agree

on the meaning of whatever signals are involved. If there is disagreement, then

communication fails. Selection will therefore favor individuals that learn the

communication system of their parents precisely over those who invent appropriate-

sounding words for concepts that they want to express (Pinker and Bloom 1990). While

this would tend to preserve sound symbolism if early language was highly sound

symbolic, children must also be able to learn arbitrary forms. Children that learn

whatever language their parents speak and pay no attention to whether words “sound

right” will be better able to communicate than children that have trouble learning

words that do not “sound right” to them. For this reason, it is possible that the

separation of language acquisition from intuitions about sound symbolism and how

words should sound would be adaptive.

It is further conceivable that the necessity for children to learn language without

potential interference from notions about what “sounds right” could occur by

separation of the language faculty as a whole from sound symbolic intuitions. This is

supported somewhat by evidence that we have different modes for listening to speech

and non-speech sounds. When we are listening to speech sounds, we generally deduce

and attend to the sequence of abstract phonemes. Most of the acoustic information that

allowed us to deduce the phonemes is either shut out or occasionally it can be

subconsciously registered (Tsur 2006). Thus sound symbolic cues (like frequency) in

the acoustic signal would seem to receive little or no attention during normal language

processing. More evidence supporting a separation between the language faculty and

sound symbolic intuition comes from evidence from a lexical decision task described

previously. Stimuli (words and non-words) were presented in spiky or curvy frames.

When the sounds composing the stimuli symbolically matched the frame, decision

times were shortened if the stimulus was a non-word, but not if it was a real word

(Westbury 2005). The interference from a frame that did not match a stimulus’ symbolic

quality would seem similar in origin to the interference that a child might experience

learning a word when the word’s meaning disagreed with the word’s symbolic quality.

Since words were processed much faster than non-words, there is some evidence that

words are processed before sound symbolic intuition can have an effect, and this would

be advantageous for children learning language.

As with most types of behavior, language is under selective pressures that favor

opposite things. Sound symbolism would be favored because it tends to make

language easier to process and more understandable. On the other hand, arbitrariness

will be favored because it is necessary for full expression, and learning language as if it

were arbitrary will tend to be the most successful language learning strategy. The

competing selective pressures favoring sound symbolism and arbitrariness, however,

may not currently be of equal strengths. It is worth noting that the incorporation of

sound symbolism into language would be most adaptive at the time language arose.

Once language and the common vocabulary were well established, there would not

seem to be as strong a selective advantage for the incorporation of sound symbolism,

and the selective pressure toward arbitrariness may have become dominant. This may

explain why there currently seems to be a separation between the language faculty and

sound symbolic intuition, and also why sound symbolism only seems to be

inconsistently incorporated into natural languages.

7 Conclusion

The body of research investigating universal sound symbolism has so far

provided equivocal evidence for its existence. Speakers of a handful of unrelated

languages share some sound-meaning associations, but there have not been enough

studies with non-English speakers to rule out that these associations are language-

specific and that agreement so far has been due to chance similarities in the languages

studied. English, Chinese, and Japanese speakers seem able to guess the translations of

pairs of antonyms in a foreign language more often than chance. From the results of

this paper it seems that at least for magnitude-related words, they are able to do so

predominantly by using the frequency code, which is a sound symbolic pattern.

However, the same pairs of adjectives which they were able to judge correctly seem to

be symbolic only slightly and not significantly more often than a standard pair of nouns

or verbs. Other studies examining the distribution of high frequency vowels among

magnitude words have also failed to find a significant correlation with diminutive

words. There does not seem to be a strong constraint to follow sound symbolic patterns

on the majority of the lexicon. On the other hand, there is evidence for a constraint on

some parts of it. Ultan’s findings suggest that ablauting is highly constrained when it

indicates size or distance. Woodworth’s findings also suggest a constraint on deictic

pronouns, place adverbs, and directional affixes. In this case it is simpler to postulate

that the constraint is universal and explain the exceptions than to explain the

independent creation of so many forms following the same pattern. This is the

strongest evidence for a universal sound symbolism to date, and the apparently

widespread existence of exceptions and counterexamples may be explained by the

selective advantages of arbitrariness in language.

References

Austerlitz, R. 1994. Finnish and Gilyak sound symbolism – the interplay between

system and history. See Hinton et al. 1994, pp. 249- 256.

Bergen, B.K. 2004. The psychological reality of phonaesthemes. Language 80. 290-311.

Brackbill, Y., Little, K.B. 1957. Factors determining the guessing of meanings of foreign

words. Journal of Abnormal and Social Psychology 54. 312-318.

Brown, R.W., Black, A. H., Horowitz, A.E. 1955, Phonetic symbolism in natural

languages, Journal of Abnormal and Social Psychology 50. 388-393.

Brown, R.W. 1958. “Phonetic symbolism and Metaphor.” In: Words and Things.

Glencoe, IL: The Free Press, pp. 110- 154.

Brown, R.W., Nuttall, R. 1959, Methods in phonetic symbolism experiments, Journal of

Abnormal and Social Psychology 59. 441-445.

Childs, G.T. 1994. African ideophones. See Hinton et al. 1994, pp. 181-194.

Davies, N.B., Halliday, T.R. 1978. Deep croaks and fighting assessment in toads Bufo

bufo. Nature 274. 683-685.

Davis, R. 1961. The fitness of names to drawings: a cross-cultural study in

Tanganyika. British Journal of Psychology 52. 259-268.

Fitch, W.T. 2000. The evolution of speech: a comparative review. TRENDS in Cognitive

Sciences 4. 258-267.

Fromkin, V., Rodman, R., Hyams, N. 2003. “What is Language?” Language. Boston, MA:

Heinle, pp. 3-12.

Hamano, S. 1998. The sound-symbolic system of Japanese. Cambridge, UK:

Cambridge University Press.

Hinton, L., Nichols, J., Ohala, J. eds. 1994. Sound symbolism. Cambridge, UK:

Cambridge University Press.

Klink, R.R. 2000. Creating brand names with meaning: the use of sound symbolism.

Marketing Letters 11. 5-20.

Maltzman, I., Morrisett, L. Brooks L.O. 1956. An investigation of phonetic symbolism.

Journal of Abnormal and Social Psychology 53. 249-251.

Marks, L.E., Hammeal, R.J., Bornstein, M.H., Smith, L.B. 1987. Perceiving similarity

and comprehending metaphor. Monographs of the Society for Research in Child

Development 52. 1-100.

Maurer, D., Pathman, T., Mondloch, C.J. 2006. The shape of boubas: sound-shape

correspondences in toddlers and adults. Developmental Science 9. 316-322.

Morton, E.S. 1994. Sound symbolism and its role in non-human vertebrate

communication. See Hinton et al. 1994, pp. 348- 362.

Newman, S. 1933. Further experiments in phonetic symbolism. American Journal of

Psychology 45. 53-75.

Nowak, M.A., Komarova, N.L. 2001. Towards and evolutionary theory of language.

TRENDS in Cognitive Sciencs 5. 288- 295.

Nuckolls, J.B. 1999. The case for sound symbolism. Annual Review of Anthropology 28.

225-252.

Ohala, J. 1994. The frequency code underlies the sond-symbolic use of voice pitch. See

Hinton et al. 1994, pp. 325- 347.

Ostwald, P.F. 1994. Some observations on the function of sound in clinical work. See

Hinton et al. 1994, pp. 309- 324.

Parault, S.J., Schwanenflugel, P.J. 2006. Sound-symbolism: a piece of the puzzle of word

learning. Journal of Psycholinguistic Research 35. 329-351.

Pinker, S., Bloom, P. 1990. Natural language and natural selection. Behavioral and Brain

Sciences 13. 707-784.

Ramachandran, V.S., Hubbard, E.M. 2001. Synaesthesia– a window into perception,

thought, and language. Journal of Consciousness Studies 8. 3-34.

Rizzolatti, G., Craighero, L. 2005. The mirror-neuron system. Annual Review of

Neuroscience 27. 169-192.

Rogers, S.K., Ross, A.S. 1975. A cross cultural test of the Maluma-Takete phenomenon.

Perception 4. 105.

Sapir, E. 1929. A study in phonetic symbolism. Journal of Experimental Psychology 12.

225-239.

Shimojo, S., Shams, L. 2001. Sensory modalities are not separate modalities: plasticity

and interactions. Current Opinion in Neurobiology 11. 505-509.

Steels, L., Kaplan, F. 2002. AIBO’s first words: the social learning of language and

meaning. Evolution of Communication 4. 3-32.

Tsur, R. 2006. Size-sound symbolism revisited. Journal of Pragmatics 38. 905-924.

Ultan, R. 1978. Size-sound symbolism. In J.H. Greenberg (Ed.), Universals of Human

Language vol. 2. (pp. 525- 568). Stanford, CA: Stanford University Press.

Westbury, C. 2005. Implicit sound symbolism in lexical access: evidence from an

interference task. Brain and Language 93. 10-19.

Woodworth, N.L. 1991. Sound symbolism in proximal and distal forms. Linguistics 29.

273-299.

Foreign Language Dictionaries

Arabic

Awde, N., Smith, K. 2004. Arabic: Practical Dictionary. New York: Hippocrene Books.

Basque

Aulestia, G., White, L. 1990. English-Basque Dictionary. Reno, NV: University of Nevada

Press.

Bemba

Hoch, E. 1960. Bemba Pocket Dictionary. Tabora, Tanganyika: TMP Printing Department.

Cayuga

Froman, F., Keye, A., Keye, L., Dyck, C. 2002. English-Cayuga, Cayuga-English Dictionary.

Toronto: University of Toronto Press.

Creek

Martin, J.B., Mauldin, M.M. 2000. A Dictionary of Creek/Muskogee. Lincoln, NE:

University of Nebraska Press.

Hausa

Newman, R.M. 1990. An English-Hausa Dictionary. New Haven, CT: Yale University

Press.

Hawaiian

Pukui, M.K., Elbert, S.H. 1986. Hawaiian Dictionary. Honolulu, HI: University of Hawaii

Press.

Hungarian

Magay, T., László, K. 1996. NTC’s Hungarian and English Dictionary. Chicago: NTC

Publishing Group.

Igbo

Echeruo, M.J.C. 1998. Igbo-English Dictionary. New Haven, CT: Yale University Press.

Japanese

Adachi, T. 1999. Kodansha’s Furigana Japanese Dictionary. Tokyo: Kodansha International.

Kikuyu

Barlow, A.R. 1975. English-Kikuyu Dictionary. T.G. Benson (Ed.). Oxford: Oxford

University Press.

Khmer

Keesee, A.P.K. 1996. An English-Spoken Khmer Dictionary. London: Kegan Paul

International.

Korean

Lee, J., Lee K. 2005. Korean: Dictionary and Phrasebook. New York: Hippocrene Books.

Malay

Winstedt, R. 1965. A Practical Modern English-Malay Dictionary. Kuala Lumpur,

Singapore: Marican and Sons.

Mongolian

Hangin, J.G. 1970. A Concise English-Mongolian Dictionary. Bloomington, IN: Indiana

University Publications.

Oneida

Michelson, K., Doxtator, M. 2002. Oneida-English, English-Oneida Dictionary. Toronto:

University of Toronto Press.

Tarahumara

Thord-Gray, I. 1955. Tarahumara-English, English-Tarahumara Dictionary. Coral Gables,

FL: University of Miami Press.

Tibetan

Goldstein, M.C. 1984. English-Tibetan Dictionary of Modern Tibetan. Berkeley, CA:

University of California Press.

Trukese

Goodenough, W.H., Sugita, H. 1980. Trukese-English Dictionary. Philadelphia, PA:

American Philosophical Society.

Turkish

Iz, F., Hony, H.C. 1992. The Oxford Turkish Dictionary. Oxford: Oxford University Press.

Zapotec

Britton, A.S. 2003. Zapotec-English, English-Zapotec. New York: Hippocrene Books.

Zulu

Doke, C.M., Malcolm, D.M., Sikakana, J.M.A. 1958. English-Zulu Dictionary.

Johannesburg, South Africa: Witwatersrand University Press.

Appendix I: Survey

Directions: This experiment tests your intuitions about the meanings of foreign words

based on how they sound. Each of the word pairs below is a translation of the English

pair “big- small” or “small-big” into a different language. For each pair, please read the

two words aloud to yourself, so you can hear how they sound, then bold/underline/

circle the one you think means “small.” Do not think about a pair for more than 5-10

seconds- go with your first intuition. If at any point during the test you feel that you

have looked at too many pairs and no longer have intuitions, please do something else

for a while and finish the test later.

Please fill in any languages that you know, have studied, or are familiar

with:____________________________________________________________

i. (Example) grande pequeño

1) udin atin

2) handi chiki

3) ncane khulu

4) saghir kabir

5) tras grish

6) jagun kuda

7) haju habi

8) iki nui

9) zench chedj

10) smar shil

11) pako bidu

12) nta agi

13) nini nene

14) inkush vadong

15) brelko egralp

16) ah ge

17) kulu nono

18) yaze yeggi

19) besar kechil

20) watla kwona

21) opor dimir

22) ibig dabo

23) waru uta

24) athek felpo

25) chooncoon chempo

26) aflet pliki

27) hugi ahob

28) kicsi nagy

29) kicik iri

30) minag ziji

31) mwuk ngin

32) kilsu balaz

33) godan ipich

34) da ob

35) cho twi

36) gujre shekri

37) fok thig

38) filka orveb

39) dondag gigank

40) babba tsigil

41) kizta shozat

Appendix II: Brown Data

Percentage of correct translations for each pair in three languages for experimental conditions A (words written and

spoken, n= 86) and B (words written only, n= 16)

% Correct % Correct % Correct

English Chinese A B Czech A B Hindi A B

1. beautiful mei 88* 70 krása 57 31 khubsurat 64* 50

ugly ch’ou osklivost badsurat

2. blunt tun 78* 70 tupy 81* 83* gothil 68* 83*

sharp k’uai spichaty tez

3. bright liang 67* 90* svetly 64* 77 chamakdar 51 90*

dark an tmavy dhundhala

4. coarse ts’u 65* 70 hruby 21* 44 mota 48 31

fine his drobny achha

5. down hsia 10* 31 dolu 56 50 niche 75* 83*

up shang nahoru upar

6. drunk tsui 66* 50 opily 21* 70 nashe men 80* 77

sober hsing strizlivy sanjida

7. dry kan 72* 70 suchy 44 50 sukha 42 44

wet shih mokry bhiga

8. fast k’uai 83* 83* richly 87* 83* tez 27* 57

slow man pomaly sust

9. fat fei 31* 57 tlusty 69* 77 mota 66* 57

thin shou tenky patala

10. gold chin 57 57 zlato 19* 57 sona 42 64

iron t’ieh zelezo loha

11. bad huai 34* 64 zly 62 57 kharab 64* 31

good hao hodny achha

12. happy huan 38 50 radostny 57 64 khush 17* 38

sad pei smutny ranjida

13. hard kang 97* 83* tvrdy 76* 96* sakht 61 64

soft jou mekky narm

14. light ch’ing 93* 90* lehky 66* 77 halka 36* 57

heavy chung tezky wazani

15. long ch’ang 55 44 dlouhy 80* 70 lamba 90* 70

short tuan kratky chhota

16. many to 73* 57 mnoho 55 25 bahut 88* 90*

one yi jeden ek

17. strong ch’iang 37 64 silny 28* 64 mazbut 34* 31

weak jo slaby kamzor

18. sweet t’ien 58 51 sladky 24* 25 mitha 88* 70

sour suan kysely khatta

19. thunder lei 23* 31 hrom 92* 96* garaj 62 77

lightning tien blesk chamak

20. warm nuan 73* 50 teply 69* 77 garam 66* 77

cool liang chladny thanda

21. wide kuan 37 90* siroky 43 57 chaura 76* 51

narrow chai uzky tang

* p ≤ 0.01