Literature Review
Theory and Review

3. Theoretical Preliminaries


3.1 Recapitulation of Basic Issues

In this dissertation, my primary purpose is to outline experiments which could in principle disprove the following generalizations, but which in fact support it:

The Phonosemantic Hypothesis
In every language of the world, every word containing a given phoneme has some specific element of meaning which is lacking in words not containing that phoneme. In this sense, we can say that every phoneme is meaning-bearing. The meaning that the phoneme bears is rooted in its articulation.

The Arbitrariness of Reference
Word semantics is not reducible to 'word reference'. The referent of a word cannot be predicted from how it sounds or how it is articulated. Words which share a common element of reference are said to fall in the same 'semantic class'. The fewer exact synonyms that a word has (the smaller the set of words that share its referent exactly) the more 'concrete' its 'reference'. The salience of iconic meaning in a word is related inversely to the concreteness of its reference.

Phonosemantic Association
When semantic domain S is associated disproportionately frequently with phoneme X, then people will be inclined to associate semantic domain S with phoneme X productively.

True Iconism
The connotation of a word is affected directly by its phonological form. On the Iconic level, a word means what it is. The form of a word does not directly affect what the word refers to, what its argument structure is, or any other aspect of its meaning. It only affects what the thing referred to in the word is like. Viewed from the perspective of parole, True Iconism or simply Iconism is among the least salient aspect(s) of word semantics. From the perspective of langue, Iconism is the most fundamental and pervasive aspect of word semantics on top of which all other layers of semantics are built.

The basic premise I maintain that allows me to assert that some aspects of every word's semantics are arbitrary and others aspects are not (i.e. that both arbitrariness and non-arbitrariness hold of every word) is this:

Word Semantics is Structured
Word semantics cannot be reduced to reference. A word's semantics is affected among other things by its part of speech, the way it fits into the semantic class structure of the language, its argument structure and its phonological form. Some of these aspects of word semantics are 'arbitrary' in nature (in Saussure's sense) and others are not.

As we have seen, there are basically two types of tests that phonosemanticists have conducted over the centuries.

1. The existing vocabulary of a given language is classified according to both phonetic form and semantic domains to see whether certain phonemes are more or less prevalent in certain semantic domains than in others. The initial 10 experiments are of this type.

2. Informants are prompted with sounds, images, foreign words or nonsense words and asked to provide some kind of feedback based on their linguistic intuitions. These results are then examined to see if there if they display any correlation between sound and meaning. The final 4 experiments are of this type.

Experiments of the first type can measure the extent of the phonosemantic influence in existing vocabulary and can provide us with a great deal of information about the nature of phonosemantic correlations. But no matter how many such experiments are run, they cannot in principle show conclusively whether or not the phonosemantic correlations are historical artifacts of an earlier linguistic process, or whether they reflect a natural law which must be completely general and which must therefore actively and synchronically affect every word in every language. To demonstrate this, one must perform experiments of type 2. Experiments of type 1 primarily provide evidence for Von Humboldt's type 3 iconism -- Clustering. And experiments of type 2 primarily provide evidence for Von Humboldt's type 2 iconism -- True Iconism.

Many of the publications in linguistic iconism in prior centuries and even in this one are more philosophical in nature than empirical. All of those which deny any regular relationship between sound and meaning that I have encountered are philosophical rather than empirical in nature. That is, they all explain on philosophical grounds why such a correlation is impossible rather than actually conducting a test to demonstrate that it in fact does not exist. In other words, in none of the 1970 entries in my full bibliography have I found an article in which an empirical test of one of the above two types was conducted with pervasively negative results. Indeed many linguists, such as Sapir and Bloomfield (and the present author) initially conducted such tests believing the results would give little or no support for linguistic iconism and ended up concluding the contrary.

I also intend in later chapters to provide an outline of the theoretical ramifications of these findings, but I first want to offer a fairly thorough corpus of supporting empirical data, for this is the foundation on which any succeeding discussions must stand.

3.2. Classification Systems

Before I continue on to discuss the tests, I would like to distinguish some different types of classificational systems. I discuss classificational systems in such detail, because the primary form of evidence I use for the Phonosemantic Hypothesis is the possibility of creating a certain kind of classificational scheme for all words which match a given phonological characterization. Consider what must be shown in order to demonstrate that a phoneme has meaning. It must be shown that all words which contain that phoneme have some element of meaning which words not containing that phoneme do not have. That is, it must be shown that words containing a given phoneme are associated with some semantic domain that words not containing that phoneme are not associated with. We can assume that if there is any such semantic domain, it will be very broad and very abstract, since there are only 24 consonants in English. So how would such a thing be demonstrated? We would first classify all the words containing a given phoneme, and then show that other words which don't contain that phoneme don't fit into that classification. This is what I propose to do, then, and this is the reason I will expend some effort now discussing the nature of classification.

In order to show that phonemes have meaning, we must have some notion of a coherent semantic domain and a coherent phonological description. I begin therefore by defining a natural set of words.

Natural Set
The set of all words which fit a given unified linguistic (phonological, morphological, syntactic or semantic) characterization.

For example, I will refer to the set of all words referring to food in all languages as a natural set. The set of all English words in a given person's vocabulary is a natural set. The set of all nouns containing the nominalizing suffix '-ment' is a natural set. The natural sets relevant to phonosemantics are those which have a unified phonological characterization, such as all monosyllables, or all monomorphemes beginning with /tr/. Sets defined by disjunctive characterizations such as 'the set of all words starting with /pl/ or referring to musical instruments' are not natural sets. Conjunctive sets such as 'the set of all words starting with /p/ and referring to a fruit' do form a natural set as I am defining it.

By a Natural Classification, I mean one having the properties 1-4 below.

Natural Classification
Criterion 1. Very nearly every word within the given natural set fits in some semantic class.
Criterion 2. Each semantic class contains a large percentage of the words in that natural set.
Criterion 3. There are relatively few semantic classes in the classification.
Criterion 4. The semantic classes in the classification are distinct

The classes formed by a Natural Classification are called Natural Classes. There is a lot of imprecision in this definition, in terms such as 'very nearly' and 'relatively'. Hopefully, further research will be able pin these down to precise ratios and percentages. For now, I appeal to the reader's common sense. An example of a Natural Classification for the words referring to foods would be:

Meats: ham, steak, beef,...
Milk Products: milk, butter, cheese,...
Fruits: peach, plum, apple, orange,...
Vegetables: potato, tomato, carrot, pepper,...
Sweets: cake, cookie, candy,...
Grains: wheat, oat, rice,...
Breads: bread, bun, muffin,...

A classification for the natural set of food which does not fit the criteria 1-4 would be:

Purple Food: plum, grape
Food Made from Petroleum:
Violet Food: plum, grape
Furry Food: kiwi, coconut
Food That Is Buried in the Ground:

These are the types of classifications that occur to us most readily for any given set of data. They are psychologically real. By means of Clustering, some phonemes in a given language may gravitate more toward some Natural Classes than others, but no Natural Class is the exclusive domain of any one phoneme.

It may seem that one could devise any number of Natural Classifications for a given set of data, but as Rosch(1973) and others have shown, this turns out not at all to be the case. Language conspires to limit the Natural Classes into which words can fall. English simply does operate in terms of, for example, words for 'food' subdivided again into 'meat', 'vegetables', 'fruits', 'breads', etc.. It does not operate in terms of 'words for objects that lean at an angle' or 'words for objects that can't easily be moved' or 'food that has been buried 4 months underground'. It doesn't even operate in terms of 'round foods' or 'soft foods', even though there are a fair number of foods which are soft and round. This means that part of the 'meaning' of the English word 'mango' is that it is classified as a fruit. That fact about 'mango' is built into English itself, and it is because of this that we can make a Natural Classification for food words which includes 'fruit' as a subset, whereas if we try to classify 'food' words according to other parameters, they do not fit the four criteria for a Natural Classification.

We cannot easily abstract away from these Natural Classifications, because they lie at the very heart of what for us distinguishes a word from a mere string of sounds. It is my contention that sound-meaning is actually more fundamental than reference or Natural Classes, but because we cannot in general stretch our mind enough to abstract away from the Natural Classes, we must work within them. I therefore look at classifications which meet even stiffer criteria than that of 1-4: those which include criteria 1-4 and then some. I will look at groupings of words which both fall within a given Natural Class and which also take on certain common semantic characteristics because of commonalities in their phonological form. These are what we call the 'phonesthemes'.

Consider the /gl/ phonesthemes mentioned in the introduction:

Reflected or Indirect Light -- glare, gleam, glim, glimmer, glint, glisten, glister, glitter, gloaming, glow
Indirect Use of the Eyes -- glance, glaze/d, glimpse, glint
Reflecting Surfaces -- glacé, glacier, glair, glare, glass, glaze, gloss

These all fall within the Natural Classes of 'light' and 'seeing' which include many words which contain neither /g/ nor /l/. 'Light', for example, is not the exclusive domain of any one consonant. 'Light' is a natural semantic domain or Natural Class. In these particular cases, however, if one accepts that Semantic Association can happen as low as the level of the phoneme, then there happens to be good evidence to suggest that the 'light' in these particular words comes from the /l/ and the indirectness comes from the /g/.

The evidence that the 'light' in these words comes from the /l/ takes the form of disproportions in semantic distribution among the phonemes. For example, apart from the abovementioned 'looking' verbs: gape, gasp, gawk and gaze, which do not directly concern light anyway, no light occurs in monosyllabic words containing /g/ but not /l/. But disproportionately many words containing /l/ and no /g/ refer to some aspect of light. And the disproportions are great.

When one looks at all the semantic domains that various phonemes favor across all the Natural Classes, one finds that they have a unified semantics that lies deeper than mere adherence to some group of Natural Classes. They are like light shone through so many prisms. One must initially consider the form of each prism as well as the nature of the light that emerges from it to determine what the original light is like. I find that the phoneme means something in its own right independently of all the classes it clusters toward. And it is this original, unified essence -- not the Clustering -- which I think of as True Iconism.

In this first experiments, I will be testing for the Phonosemantic Hypothesis by trying to form a more narrow Natural Classification for various natural sets which are phonologically defined. I will call such classifications 'Phonosemantic'. Phonosemantic classifications are essentially classifications of phonesthemes. The first four criteria for a Phonosemantic Classification are merely the criteria for a Natural Classification as defined over natural sets which are phonologically defined:

Phonosemantic Classification
Criterion 1. Very nearly every word with the given phonological characterization fits in some semantic class.
Criterion 2. Each semantic class contains a large percentage of the words which match that phonological characterization.
Criterion 3. There are relatively few semantic classes in the classification.
Criterion 4. The semantic classes in the classification are distinct
Criterion 5. Each word fits into an average of a fairly large number of classes.
Criterion 6. The semantic classes are narrowly defined. By a 'narrowly defined' semantic class, I mean one which encompasses a small percentage of words in the language as a whole.
Criterion 7. A much smaller percentage of the words which do not match the relevant phonological characterization fit into any class.
Criterion 8. Those words that do not match the relevant phonological characterization but which nevertheless do fit in the classification fit on average in a smaller percentage of classes, than those words which do match the phonological characterization.
Criterion 9. Any class in a Phonosemantic Classification can be defined narrowly enough that words not matching the relevant phonological characterization are excluded from it.

This is a very tall order to fill, but in my view, these are the criteria that must hold if phonemes are to be shown to be meaning-bearing. And in my view, these criteria do indeed hold English. Before going on to more detailed tests, I will provide here a small illustrative example of the type of data that concerns us.

3.3. A Small Scale Example of the Phonosemantic Experiment

In attempt not to lose the forest for the trees in our discussion, let me now give a brief overview of the types of tests which will be conducted on a much larger scale in the following chapter. Consider once again /gl/ in initial position. (The upcoming tests will hopefully convince the reader that any polyconsonantal onset works equally well, but for the purposes of exposition, I find it clearer not to keep presenting new data.) One possible Phonosemantic Classification for English monomorphemes beginning with /gl/ might look like this:

Reflected or Indirect Light -- glare, gleam, glim, glimmer, glint, glisten, glister, glitter, gloaming, glow
Indirect Use of the Eyes -- glance, glaze/d, glimpse, glint
Reflecting Surfaces -- glacé, glacier, glair, glare, glass, glaze, gloss
Other Light or Sight -- globe, glower
Understanding -- glean, glib, glimmer, glimpse
Symbols -- gloss, glyph
Ease -- glib, glide, glitter, gloss
Slip -- glide, glissade
Quantities -- glob, globe, glut
Acquisition/Stickiness -- glean, glimmer, glue, gluten, glutton
Strike -- glance
Containers -- gland, glove
Joy -- glad, glee, gloat, glory, glow
Unhappiness -- gloom, glower, glum
Natural Feature -- glade, glen

One observes several things initially:

* The large majority of these various classes are ordinary cross-phonemic Natural Classes (Light, Sight, Surfaces, Thinking, Symbols, Motion, Quantity, Acquisition, Strike, Containers, Joy, Sorrow, Natural Features). One finds light, understanding, symbols, etc. in many other consonant sequences besides /gl/. For one thing, one finds reflected light in the word 'gleam', 'glim', 'gloaming' and 'glimmer', which all contain an /m/. These phonesthemes at first glance represent merely semantic disproportions among phonemes. And we cannot even be sure that they are disproportions unless we try sorting all other phonologically defined Natural Sets into these same classes. What one really sees in Phonosemantic Classifications of this type is the way that /gl/ manifests through the filters of these various Natural Classes. It remains to be seen whether /str/ or /fr/ pattern any differently.

* These classes are related to one another. There is, for example, a quite general thematic metaphor in English "Light IS Understanding" (Lakoff and Johnson, 1980). The symbols in /gl/ (gloss and glyph) are also related to understanding. Similarly, acquisition is related to sight and to quantities (glob, glut) and to containers. This preoccupation with acquisition, quantities and containers is quite general to the velar consonants. Ease, joy and understanding are also related to one another. This interrelatedness of the most prevalent semantic domains for a given phoneme is one of the first intimations that each phoneme and phoneme sequence actually has a unified meaning which underlies all of these classes.

* One can also notice here that in addition to single consonants, often multiple consonant combinations will confine themselves to a semantic domain which is narrower than the sum of the parts requires. All of the semantic domains listed here are theoretically available within the semantic confines imposed by /g/ and /l/. But /gl/ tends to like to confine itself even more than necessary so that 40% of these /gl/ words concern light. This patterning in /gl/ is typical in that many of the natural semantic domains which are possible for /gl/ are in fact represented in English, but not uniformly so. Once again, this is an example of 'Clustering' or Phonosemantic Association.

* This clustering is to some degree specific to English. A high percentage of the words beginning in /gl/ in all the Germanic languages concern reflected light. In Russian, for example, too, there is a certain amount of vision (gladet' -- gaze, glanut' -- cast a glance, glaz -- eye, glazet' -- stare) and a lot of smooth surfaces (gladit' -- iron, gladkij -- smooth, glad' -- mirror-like surface, glazirovat' -- glaze, glazur' -- icing, glissir -- hydroplane, glyanets -- polish) but the percentage is lower due to the fact that other basic words begin with /gl/ and form vorteces for Clustering (glava -- head, main, glubok -- deep, golos' -- voice, glina -- clay, glup -- stupid). Clustering tends to be more language-specific, whereas true Iconism is universal.

Classifications like the one for /gl/ above verify criteria 1-6 for a Phonosemantic Classificational scheme. But in order to check for criteria 7-9, we must try putting words with a different phonological characterization into the classes tailored for /gl/. I will provide here one small example of this in order to demonstrate what happens quite generally cross-linguistically and cross-phonemically. Look first at the /fr/ words that do fit the characterizations provided for the /gl/ phonesthemes.

Reflected or Indirect Light -- glare, gleam, glim, glimmer, glint, glisten, glister, glitter, gloaming, glow
Indirect Use of the Eyes -- glance, glaze/d, glimpse, glint
Reflecting Surfaces -- glacé, glacier, glair, glare, glass, glaze, gloss
Other Light or Sight -- globe, glower: fresco
Understanding -- glean, glib, glimmer, glimpse: frame (a question)
Symbols -- gloss, glyph: franc
Ease -- glib, glide, glitter, gloss: frank, free
Slip -- glide, glissade
Quantities -- glob, globe, glut: fraught, freight
Acquisition/Stickiness -- glean, glue, gluten, glutton: fraud, free, freeze, frisk, frog
Joy -- glad, glee, gloat, glory, glow: frank, free, frisk, frivol
Unhappiness -- gloom, glower, glum: fray, frazzle, fret, fright, frown
Natural Feature -- glade, glen
Containers -- gland, glove: frame, fret, fridge, frieze, frill, fringe
Strike/Touch -- glance: fray, french, frisk
Exceptions -- frail, frappé, freak, freckle, frenzy, fresh, friar, friend, fritter, frizz, frizzle, frock, from, frond, front, frontier, frosh, frost, froth, frowzy, fruit, fry

Obviously, a much smaller percentage of /fr/ words fit in this classificational scheme than /gl/ words, and those words that do fit appear a little less frequently on average (an average of 1.3 times for /gl/ vs. 1.1 for /fr/). Thus criteria 7 and 8 of the Phonosemantic Classification are met in this small example. Furthermore, observe that in all these cases, the /fr/ words that fit the characterization given for these /gl/ phonesthemes actually fit the characterization differently (criterion 9). Whereas 41% of words beginning with /gl/ concern sight, only one /fr/ word 'fresco' is marginally related to sight. The understanding in /gl/ is receptive. In /fr/, the one word which marginally concerns understanding (frame a question) is directed outward toward the source of information rather than inward toward the one who understands. The 'Ease' in /gl/ concerns non-difficulty or superficiality. The 'Ease' words in /fr/, 'frank' and 'free', are oriented rather toward openness and liberty. The quantities in /fr/ (freight and fraught) both imply a predication -- something which is fraught or weighed down with or by something else. This is not true of the corresponding /gl/ words. The unhappiness in /gl/ is depressed. In /fr/ it is largely nervous or afraid. The /fr/ words classified here as containers are not really containers at all, but frames, borders or edges. The joyfulness in /gl/ differs from that in /fr/ in that it is more inwardly than outwardly expressed.

It is typical that in semantic comparisons of two different phonologically defined classes of words that half the words in each group don't fit in the other group's semantically based classification at all. The /fr/ words taken as a whole fall more easily into a different Natural Classification. Overall, when one seeks out a Phonosemantic Classification for /fr/, one gets a very different profile than what one finds in /gl/, even though, once again, most of the classes are natural and therefore not limited to a particular phonological form. Words beginning with /fr/ which did fit in the above /gl/ scheme are italicized:

Vulnerable, Young -- frail, freak, fresh, frosh, fruit, fry
Deceit -- frame, fraud
Freeze, Congeal -- frame, frappé, freeze, fresco, fridge, frost
Frame, Border -- frame, fret, frieze, frill, fringe
Disintegrate -- fray (come apart), frazzle, free, fritter, frizzle, frowzy, fry
Fuzzy, Frilly -- frappé, fray (come apart), frazzle, freckle, frieze, frill, fringe, frizz, frizzle, frock, frond, frost, froth
Nervousness, Fear -- fray, frazzle, frenzy, fret, fright, frown
Front, Far -- from, front, frontier
Friendly People -- friar, friend
Burden -- fraught, freight
Fun and Free -- frank, free, frisk, frivol
Exceptions -- franc, frog

The English /gl/ and /fr/ words, then, do meet the criteria of the Phonosemantic Classification, and therefore the sounds /gl/ and /fr/ appear to be affecting the meanings of the words that contain them.

An important question I ask in the following experiments, then, might be phrased as this: For any arbitrary phonological characterization in any arbitrary language, can a classification be found which meets criteria 1-9 for a Phonosemantic Classification above? That is, is this just a historical artifact peculiar to English, or is some active, productive natural law at work? The initial tests in this chapter concern classifications of this type, many of them on a fairly large scale. I believe that in the experiments I outline below, I have found classifications which taken as a whole do meet these 9 criteria for a Phonosemantic Classification for the monosyllabic vocabulary of English. To the extent that these tests yield the same results in languages and semantic domains not covered in this work, then I cannot see but that it must be admitted that phonological form significantly affects the semantics of words universally.


After spending some time at this, the big question that begins to loom over one takes the form of, "What after all is the semantic distinction between 'gleam' and 'glimmer' and 'glisten' and 'glitter' and 'glow'? How do we characterize it? How do we learn it? Where does it come from?" It seems to me that the results of these experiments suggest that the essential differences between the words which fall together in narrow natural semantic domains of this nature can be attributed to the effects that phonology has on semantics. Once a phoneme is filtered through the Natural Classes, its semantic effect becomes amazingly specific. Consider once again the famous reflected light phonestheme this time sorted according to the other sounds which occur in these words:

Reflected or Indirect light


gleam, glim, glimmer, gloaming

glint, glitter

glisten, glister



glimmer, glister, glitter





gloaming, glow

glim, glimmer, glint, glisten, glister, glitter

The most intense of these words is clearly 'glare'. Furthermore, /r/ occurs proportionally more frequently than any other consonant in words associated with 'intensity' in every natural semantic domain. (Genette has devoted a chapter to this, and my findings confirm it.) The /gl/ words which refer to a sparkly kind of light all contain a /t/. Those that additionally contain an /s/ are more intense and less superficial. This intensification is quite pervasive in English words containing /s/. (Consider effects of the type mash/smash, tamp/stamp, etc. ) The word 'glint' connotes a mere suggestion of light. This is quite common in words containing an /n/ in pre-final position: hint, tint, faint, point, scent, taint, scant,... also fringe, glance, pinch, strand, tinge, twinge, hunch, sound out, get wind of, etc. There's a different quality to the light in words containing /m/ than in words containing /t/. It is less sharp and sparkly. The reflection seems to be against a smoother surface. This holds of other /m/ words concerning light as well (flame and beam). The phoneme /m/'s light is also not as abrasive as the light in 'glare'. The phoneme /r/ quite generally has a 'tearing' or 'ripping' quality. It frequently occurs in words in which the integrity of form is violated. The /m/ words differ from 'glow' not so much in the quality of the light, but in the fact that the /m/ words imply a dawning or a beginning (the project was but a gleam/*glow in his eye), whereas 'glow' implies light in the fullness of its manifestation (She was all aglow/*agleam.). The 'gleam' in /m/ differs from the 'glint' in /n/ in that the /m/ suggests the beginning of something ongoing, and the /n/ suggests a hint of something caught in mid-stream. For example, if one walks past a door which is cracked open and sees a flash of light as one passes, that might be a 'glint' but not a 'gleam'. If one is motionless relative to the light but sees a touch of it through a crack that suggests that something interesting might be going on inside, then that's a 'gleam' more than a 'glint'. The labials quite generally appear in words concerning beginnings, and the dentals quite generally occur in words concerning linearity and ongoing processes.

Consider now other aspects of the phonology of these words. The disyllables that end in -er or -en all suggest a repetitive or unsteady quality to the light. This is not true of the monosyllables. The -er words imply that the sparkly effect happens all the time. The -en suggests that there is a particular light source relative to which the reflected light occurs (It's glittering, glimmering, glistering/ ?It's glistening. We prefer: It's glistening in the light of the sun.) The words containing a short 'i' all refer to light that is short-lived. The words containing other vowels all refer to light that is prolonged or ongoing. Of these, the high vowel (gleam) suggests a narrow band of light. Those containing /ow/ concern light that is not directed, all-pervasive.

It might be of interest now to consider the Norwegian words in the same semantic domain. To make the comparison with English easier, I will not include verbs of seeing or reflecting surfaces (glass, glatt, glette, glire, glitte, glitter, glor)



glans, glinse

glime, glimmer, glimte

glimte, glitre

glo, glø

Once again, the words containing a short /i/ refer to short-lived manifestations of light and those which contain other vowels in stressed position suggest more prolonged light. Once again, the word containing /r/ has an intensity the others don't. Once again, the /m/ words imply a less sparkly, smoother light than /t/, and also suggest the beginning of something. Once again, the words ending in -er/re suggest repetitiveness or an intermittent quality. Once again, the verbs that end in vowel suggest light or heat in the fullness of its manifestation. Norwegian also has a class of /ns/ words which do not occur in English words for light. These words have a quality of ease that one finds also in English 'dance', 'prance', 'glance', 'rinse', 'prince/ly', etc. The data in Appendix I suggests that the finesse and ease in these words is provided by the /n/, and the strength by the /s/, the light by the /l/ and the indirectness by the /g/. A close look at vowel semantics suggests that the short-lived quality of 'glinse' vs. 'glans' is attributable to the short /i/. Whether or not these similarities are attributable to common etymologies between the two languages, the fact remains that over thousands of years, /m/ and /t/ and /r/ all still correlate with a fairly specific and consistent aspect of light in words of both languages.

I am well aware that a discussion of this type does not constitute proof that these aspects of the phonemes are indeed affecting the semantics of /gl/ 'light' words as I suggest. But I do believe that a close look at all the data presented in Appendix I taken as a whole does constitute proof that the effect of phonology on the semantics of English monosyllables is just that pervasive and just that specific. It has been presumed that correlations of this kind are coincidental and sporadic. The data presented here and in some other works in phonosemics shows that correlations of this nature are universal and productive. They cannot therefore be coincidental. Experiment I is the largest scale experiment in this dissertation. Unless one actually works laboriously through the data in Appendix I, however, it is very hard to see many aspects of what it shows. I will therefore augment that data with other experiments which address issues which can and should be brought into question.

3.4. Overview of the Experiments Conducted

First a series of several tests were conducted which analyzed the existing vocabulary of English, and of certain subdomains of the vocabulary of languages other than English. Following that, another series of tests were conducted, all of which queried informants regarding their intuitions about the semantics of nonsense words.

What follows is a brief description of the tests that I have run and will discuss below in more detail throughout the dissertation. The actual data resulting from each of the tests can be found in Appendices I-XIV.

Experiments Which Analyze Existing Vocabulary

Experiment 1 -- Classification First by Phoneme Sequence then by Semantic Domain
(Section 4.1, Full data and results in Appendix I)

In this experiment, I extracted all the monosyllables familiar to me from Houghton Mifflin's American Heritage Dictionary. These words were divided into 24 classes, one for each consonant phoneme. I then devised a tentative phonosemantic working classification for each of these subclasses. Finally, the words within each of the resulting phonesthemes were subdivided again according to the phoneme's position within the syllable.

Only 3% of the 3485 monosyllables did not fall easily into a Phonosemantic Classification. All of these exceptional words did, however, fall into a limited set of Concrete Noun classes, that is to say they are nouns with rigid referential meanings. For example, the exceptional 'body parts' were 'beak, jowl, thigh'. The exceptional games were 'craps', 'golf' and 'whist'. To some extent, a different Phonosemantic Classification results in a different list of exceptions, but whenever I have formed a Phonosemantic Classification, all of the words which don't conform to the classification end up being Concrete Nouns. In addition to these 3% that don't fit in my phonosemantic classes, there are hundreds of words that fit in both the phonosemantic classes and the Concrete Noun classes. The exceptions tend to have a single narrow and well-defined non-idiomatic function in the language. Polysyllabic monomorphemes are considerably less likely to fit in the Phonosemantic Classification than are monosyllables.

This experiment provides us with a general idea of the preferred semantic domains for each phoneme and the percentages of words containing a given phoneme that can be characterized by these semantic domains. It also allows us to observe the semantic effect of phoneme position. Finally, it allows us to observe the relative nature of those words which do and do not easily submit to Phonosemantic Classification, namely that they are Concrete Nouns.

Precedents in the Literature -- There have been a number of studies which perform phonosemantic analyses of existing vocabularies.4 Unlike most of those predecessors, the present experiment lists not only words which fall into the phonesthemes listed, but also those which do not fall into phonesthemes listed. This, I believe, is significant, for only by covering the entirety of a well-defined portion of a vocabulary is one able to quantify the extent of the disproportions. And unless one can quantify the extent of the disproportions, an interesting hypothesis has perhaps been presented, but nothing substantive has been proven.

I've found only a very few works which cover large portions of a language's vocabulary in its entirety, none of them on as large a scale as my first experiment here. The most notable precedents that I have come across are Bloomfield(1910), Ertel(1972), McCune(1983) and Lawler(1990). Neither Bloomfield, Lawler nor McCune take statistics in the manner that I do in Experiment 1, showing precisely what percentage of the vocabulary with a given phonological form falls within each phonestheme. Their works do, however, sketch the most important results of such statistical analyses, since they do cover the entirety of a well-defined portion of a vocabulary for a language or language group. Ertel does provide statistics over his results, but his experiment is more similar to later experiments I conducted, so his results are not commensurable with those of this experiment.

I believe this Experiment 1 is the first attempt to provide a semantic profile of individual phonemes in a systematic way over a large range of words. McCune mentioned in his dissertation that he thought it possible that even phonemes could be shown to have meanings and Richard Rhodes (personal correspondence) has told me that he independently realized that this was the case some years ago, but he has not had an opportunity to write anything up about it. Though one finds occasional mention of positional effects in the literature, I believe I am also the first to undertake in a systematic way an analysis of the positional effects of individual phonemes on meaning, as I do implicitly here in the first test and explicitly in the 8th and 9th tests (sections 4.8 and 4.9).

Experiment 2 -- Classification First by Phoneme Sequence, Subclassification by Semantic Domain and then Regrouping of Different Phonemes by Semantic Domain
(Section 4.2, Full data and results in Appendix II)

In this experiment, all the monomorphemes in my active English vocabulary containing an /r/ in second position were classified by initial consonant. Then for each of these subclasses, a Phonosemantic Classification was created. Then the phonesthemes in each of these groups for similar Natural Classes were matched up. For example, the 'breaking' phonestheme for /br/ was aligned with the 'fracturing' phonestheme for /fr/, and an attempt was made to determine how these matching phonesthemes differed semantically. I find that one sees the effects of individual phonetic features more clearly if one conducts the experiment this way rather than by finding all the 'breaking' and 'fracturing' words first and then subdividing by initial consonant.

This experiment gives us a better view of what specific role each phoneme plays within a given semantic domain than the previous experiment. However, since it confines itself to a more limited portion of the English vocabulary, it does not so readily give one a broad overview of the semantics of each phoneme as did the first experiment. The level of specificity also allows one to see more clearly what the phonetic features have in common semantically. The semantics of the phonetic features is, of course, even more abstract than that of each of the phonemes.

Precedents in the Literature -- I don't know of any tests in the literature which fit this description. I have found this particular type of experiment the most effective and reliable means of identifying a semantic characterization of phonetic features. Many papers contain brief notes about the apparent semantics of particular phonetic features, but I've seen no attempt to conduct experiments which tried to get at these meanings in any systematic way. Wescott's (1971) paper on labiovelarity and derogation comes to some similar conclusions, though his methods are quite different.

Experiment 3 -- Natural Classes for Arbitrary Sets of Words
(Section 4.3, Full data and results in Appendix III)

In the first two experiments, one analyzes sets of words into a Phonosemantic Classifications. As one conducts these experiments, the big question that looms in one's mind is the extent to which all other words would fit just as easily into these same classes. Subjectively it may seem impossible that any arbitrary class of words could fit equally well into these same classes -- that 40% of all the words in the language could refer to reflected light, for example -- but often the judgements are subtler, and one can fool oneself. This test is a reality check. In it, each 10th word in alphabetical order is classified into a Natural Classification. The result is then compared with the Phonosemantic Classification found in Experiment 1for words beginning with /b/.

One finds that words chosen at random do fall into a limited number of classes -- the Natural Classes -- but that these classes are neither as limited nor as specific as those which words beginning with /b/ can be classified into. Furthermore, once Natural Classes for a random set of words are formed, one finds that some classes have a preference for certain phonemes over others. In comparing this classification with a scheme found for words beginning with /b/, one finds certain of these classes represented in the /b/ words in large quantities and about half completely devoid of words beginning with /b/.

Precedents in the Literature -- I don't know of any papers comparing Phonosemantic Classifications for a given phonological form with random words.

Experiment 4 -- Classify Words Containing a Phoneme Sequence X into a Classification Designed for Words Containing Phoneme Sequence Y
(Section 4.4, Full data and results in Appendix IV)

Another way to check whether all phonemes can be classified alike is to take a classification that works very well for words containing one phoneme and try to fit words containing another phoneme into these same classes as I did in the illustrative example comparing /gl/ and /fr/. In this case, I took all the monosyllables containing /l/ and tried to fit them into the /b/ classes. I found that whereas I could count on one hand the number of /b/ words which didn't fit in this scheme, about half of the words containing /l/ didn't fit in the /b/ classes at all. Those that fit best were those that also contained a /b/ or /p/ -- the class of 'bulging' and 'roundness'. Furthermore, those that did fit, fit differently. That is, although the defining (natural) characterization given for the /b/ class also fits some of the /l/ words, the words containing /l/ nevertheless differed from those containing /b/ in some observable way.

Precedents in the Literature -- I don't know of any published experiment of this type.

Classification First by Semantic Domain, then by Phoneme

These first four tests all begin with sets of words which have some common phonological characteristic and then classify them semantically. In the following tests, the procedure is reversed. I begin the next few tests by choosing all the words which fall in a natural semantic domain. I then subdivide these words according to their phonological form to see of I can discern any patterns. In this way, we can get a better overview over the phonosemantics of individual Natural Classes as well as a little more insight into Iconism (as opposed to only Clustering).

Experiment 5 -- Words Referring to Walking
(Section 4.5, Full data and results in Appendix V)

In this experiment, I looked at all the monosyllabic verbs of motion in my English vocabulary which must be done on foot. These included primarily verbs of walking, running, jumping and dancing. One finds in this case that each phoneme has a surprisingly specific effect on the meaning in the context of a sufficiently narrow class. For example, all such walking verbs which begin with a /t/ imply that the walking has a specific goal which may or may not be reached. Those containing /p/ imply that the walking involves discrete steps, and so forth.

Precedents in the Literature -- This is one of the more common types of phonosemantic experiment and it finds many precedents in the literature. There are several studies which subdivide words first into semantic domain and then analyze their phonological form in a manner similar to what I present in tests 5 and 7. McCune discusses the phonetics of various semantic domains in Indonesian in a similar manner. Such papers also include André (1966), Barry and Harper (1995), Bolinger (1946), Callebaut (1985), Cassidy, Kelly & Sharoni (2000), Chan (1995), Emeneau (1938), Ertel(1972), Fónagy (1963), Fónagy and Fónagy (1970), Gordon and Heath (1998), Greenberg and Sapir (1978), Hines (1994), Hough (2000), Jurafsky (1996), Langdon (1994), Leman (1984), Lihomanova (1999), Pentland (1975), Prokofieva (1995), Rhodes (1980, 1981, 1994), Sapir (1911), Tanz (1971), Traunmüller (1996), Wescott (1971, 1975, 1977, 1978), Whissel (1998), Whissel and Chellew(1994), Woodworth (1991). I'm not aware of any studies specifically concerning walking, though there's a study by Kendon (1972) on body motion and speech. Robin Allot has developed his phonosemantic motor theory of language based on the gestural equivalents to speech forms, and others have pointed out that phoneme meanings seem to be rooted in articulation -- the meaning of a phoneme, in other words, seems to be related to its literal physical shape. Works such as Rhodes (1994) and Wescott (1971) analyze the phonosemantic structure of words to a similar degree of specificity.

Experiment 6 -- The Bias in the Labials
(Section 4.6, Full data and results in Appendix VI)

In this experiment, certain classes were chosen which were known to favor labial consonants. These were:

Bulges, Mountains, Humps and Peaks
Fountains and Blowing
Pairs, Names, Pictures, Symbols

This experiment verifies that these classes do indeed overwhelmingly favor labial consonants. Furthermore, we find as in the previous experiment, that within such limited semantic domains, individual consonants do seem to have quite specific semantic effects.

Precedents in the Literature -- Though there are many papers which discuss the phonosemantics of a given semantic domain, I haven't found any which look at a range of semantic domains which are known to contain words characterized by a given phonological form. Emerson (1996) has done a quite thorough study of explosive words containing nasal stops, and Wescott (1971) is also similar. Neither of these is as thorough, I feel. They do not, for one thing, classify all the in the semantic domain words which do not have the relevant phonological form and they therefore cannot take statistics. However, Ertel's (1972) methodology and method of taking statistics seems to me very similar to this one. His study is cross-linguistic, and he finds the correlations between sound and meaning in the four semantic domains he researches to be universal.

Experiment 7 -- Locations
(Section 4.7, Full data and results in Appendix VII)

In this experiment, words which refer to places and which begin with certain specific consonants are taken into consideration. First the 'location' words beginning with /b/ were classified in two ways -- one in a phonosemantic manner which favored /b/ words, and the other into Natural Classes which are equally applicable to words of any phonological form. Then words beginning with consonants other than /b/ were classified first according to the one scheme and then the other. It was demonstrated in this manner that certain Natural Classifications do indeed favor a given phonological form and others do not. Finally the same experiment was performed with words for locations in Russian beginning with /b/ and also with other consonants. It was found that the Russian words patterned very much like the English words, even though the classifications were initially created only for English words.

Precedents in the Literature -- This again is similar to Ertel (1972) in that it both limits itself to a given semantic domain and then performs a cross-linguistic analysis. This experiment is on a smaller scale than Ertel's. It also differs from Ertel in two ways. First, I combine a cross-linguistic analysis of a semantic domain with the methods used in Experiment 4. And second, I show that there are two types of natural classifications for a phonologically defined natural set of words.

The Effect of Position within the Syllable

If phonosemantics is truly Iconic in nature, then every distinction in form should give rise to some kind of semantic distinction. Thus I look also for the effect of position within the syllable on word semantics. I find that these experiments can only be effectively conducted after one has convinced oneself that each phoneme does indeed command a unique semantic domain, and only after one has a sense for what the specific semantics of each phoneme consists in. In addition to the semantic effect of Natural Classification, the effect of the phoneme itself must be filtered out before one can observe positional effects. The following two tests propose ways of getting a sense for the effects of position on word semantics.

Experiment 8 -- Positional Iconism -- Comparison of Similar Phonemes
(Section 4.8, Full data and results in Appendix VIII)

In this experiment, all the English monosyllables in my vocabulary which contain /l/ or /r/ and which fall in one of the following semantic classes were classified into phonesthemes:

Non-Vehicular Motion, Vehicular Motion, Liquid in Motion, Sound, Speech, Make Active, Scare /r/ -- Calm, Slow Down /l/, Curse or Criticize, Roads

In the previous experiments /l/ and /r/ have been shown to be quite similar phonemes also semantically. Furthermore, they both appear in many positions within an English syllable. In this experiment, words containing /r/ in each position and referring to e.g. non-vehicular motion were compared with words containing /l/ in the same semantic class and appearing in the same position within the syllable. It was found that although /r/ and /l/ each have their own unified semantics, the effect of that semantics is also colored by the position that phoneme occupies within the syllable. It was also found in this case and in the previous experiments that /l/ and /r/ consistently differ from one another semantically.

Precedents in the Literature -- I've not encountered any studies which match this description. This is the best method I have found for getting at the specific semantic difference between two phonemes. The comparison between the phonemes also makes the positional effects much clearer.

Experiment 9 -- Reverse Phoneme Order
(Section 4.9, Full data and results in Appendix IX)

In this experiment most of the monosyllables in my vocabulary were taken into consideration. All words containing a given pair of consonants were classified into Natural Classes. Then all words containing the same two consonants in the reversed order were classifed into Natural Classes. Then the two classifications were compared in order to ascertain: 1. which classifications favored which order, and 2. what effect the order had on words in the same semantic class. In many cases the semantic effects of this reversal are not immediately obvious. For this reason, a detailed discussion of all the monosyllables containing /t/ and /r/ has been included in the discussion of this test.

Precedents in the Literature -- I'm not aware of any studies which match this description either.

Experiment 10 -- The Universality of Phonosemantics, the Case of /s//t//r/
(Section 4.10, Full data and results in Appendix X)

A final area which will concern us in these experiments which are based on existing vocabulary will be the universality of phoneme semantics. If the Phonosemantic Hypothesis has some validity, then it may or may not be the case that at least some of the association between phonology and word semantics is truly Iconic as opposed to being a by-product of Clustering. It may or may not be the case that these correlations are subject to natural laws and productive in living language as opposed to by-products of earlier historical processes. Indeed, though there is much indirect evidence to suggest that phonosemantics involves both Iconism and Clustering, and that it is productive, all these experiments actually only provide conclusive evidence for a conventional, non-productive as opposed to natural association between sound and meaning. To the extent that the association is in fact truly Iconic (natural), it must also be universal. In this experiment, we find that all the roots which appear in a wide variety of languages and which contain /s/, /t/ and /r/ in that order fall universally within a quite limited classification, and that words with other phonological forms do not fit in these classes.

Precedents in the Literature -- There are several studies of this type. The most prominent of these to my mind is again Bloomfield's monograph. I have also read Dempwolff (1925) and been influenced considerably by Salisbury's (1992) excellent unpublished cross-linguistic study of the k-v-n sequence. Of these, Salisbury's work is most similar to what I do here in Experiment 10.

Experiments Which Analyze Nonsense Words

In this final series of tests, I use informants' intuitions to analyze the semantics of nonsense words. These experiments allow us to examine to what extent Iconism and Phonosemantic Association are synchronically productive processes in a way that analysis of existing vocabulary never can, no matter how general the results of that analysis proves to be.

Experiment 11 -- Invented Definitions for Nonsense Words
(Section 4.11, Full data and results in Appendix XI)

In this experiment, visitors to a phonosemantic Web site were asked to provide definitions for nonsense words, such as 'baff', 'drulk', 'leb' and 'wentle'. This experiment overwhelmingly confirmed the existence and productive nature of Semantic Association, and it was argued that this Semantic Association must be happening on the level of the phoneme and not merely on the level of the word. No limitation was set on the definition that informants could provide, and yet on average 80% of the definitions fell into a few (on average about 4) fairly narrow natural semantic classes for each word, though these classes, of course, varied from word to word. Furthermore almost all of the definitions provided were very similar to the primary sense provided in the dictionary for some other common word which closely resembled the test word. It was found, however, that certain words similar to the test word were consistently imitated, and other equally similar words were not.

Precedents in the Literature -- I've been informed of the existence of one paper by Cynthia Whissel and H. Nicholson (1991) on children's invented definitions for words. I've not had access to this paper, though in general I've found Whissel's work to be very thorough and interesting. Otherwise I know of no papers that discuss the results of a experiment similar to my experiment 11. The Internet has, however, made it much easier and less expensive to collect this type of data, so I would anticipate many more tests of this type to appear in the near future.

Experiment 12 -- More Narrowly Limited Semantic Characterizations of Nonsense Words
(Section 4.12, Full data and results in Appendix XII)

In this experiment, informants were asked to define 'nem', 'forp', and 'woat' within more narrow semantic domains. Questions took the form, "If 'nem' were a size, what size would it be.?" Once again it was found that the responses were not arbitrary.

Precedents in the Literature -- The most famous work along these lines is Sapir's (1929) study in which he asks informants questions of the form, "'Mal' and 'mil' are both words for 'table' in some language. Which table is bigger -- 'mal' or 'mil'?" Most of the studies which involve informant queries (see again endnote 2) ask them to guess the meanings of foreign words within certain limitations. Such experiments are in some ways similar to my experiment 12.

Experiment 13 -- Invented Words for a Given Definition
(Section 4.13, Full data and results in Appendix XIII)

In this experiment, the order of the questioning was reversed. Informants were provided with a definition and asked to provide a word which fit this definition. The definitions used were:

to scrape the black stuff off overdone toast
to drag something heavy into the water
to swarm over the head like mosquitoes
the texture of a hedgehog
the feeling you get falling downward on a roller coaster
the appearance of the sky before a storm
a paper cutter
a layer of pollen on plant leaves
the knobs on the spikes of a hairbrush

It was found that informants strongly preferred certain phonemes over others for each of the definitions. In some cases, the phonemes which were preferred appeared in the definition itself, but this was often not the case. Indeed, in 4 cases out of 325, two informants chose one and the same nonsense word for a given definition. If one figures the number of possible English syllables to be around 50,000, then the likelihood of this occurring is about 1 in 15.

Precedents in the Literature -- Much of the commercial interest in sound symbolism has come out of the naming industry, and consequently there are some studies of applicability of names both to people and products. This is similar to what I am trying to get at in Experiment 13, though I know of no one who has queried many informants asking them to invent words for definitions, as opposed to commercial products.

Experiment 14 -- Invented Words to Describe Images
(Section 4.14, Full data and results in Appendix XIV)

It was thought in the previous experiment that the word which appeared in the definition may have predisposed informants to choose certain phonemes over others by Phonosemantic Association. Of course, if the sign is arbitrary, then Phonosemantic Association should not be a factor in their decisions. But assuming it is in some degree not arbitrary, to what extent is the association natural, and to what extent conventional? Another test was therefore devised in which the informants were prompted with semi-abstract images rather than definitions. In the previous test, informants were asked to restrict themselves to monosyllables. In this test no such limitation was imposed. It was found once again that the nonsense words provided to describe these images were far from random. Indeed, of the 204 total responses to all 6 images, there were once again 2 identical pairs, and numerous near pairs. If I take the polysyllabic and words with ungrammatical syllable structure out of consideration, then the chance that there will be 2 identical pairs among the monosyllabic responses is about 1 in 14. However, taking into consideration that many responses were disyllables, the chance of 2 identical pairs shoots way up to about 1 in 74,000.

Precedents in the Literature -- There are several studies which discuss images and sound. Davis (1961) actually conducts a similar experiment. Related works include also Berghaus (1986), Helson (1933), Khatena (1969) and Schaefer (1970).


I have organized my presentation of each experiment in a manner which I hope will not leave the reader questioning why I have drawn the conclusions that I have. I begin with a description of the methodology of the experiment followed by a concrete illustrative example. Finally I discuss the results. The complete collection of data I compiled for the experiment can be found in the appendix with the same number. For example, the data for experiment 8 is in Appendix VIII.