6. DIACHRONIC PSYCHOLINGUISTICS

In this section we discuss a variety of topics, all of which have in common the fact that they involve comparison between two or more stages in language development. Attention is first directed toward development of language behavior in the individual member of a speech community, first language learning; a general theoretical model of the process is described, a possible experimental analogue is suggested, and various research problems are discussed. A second topic is second language learning and bilingualism. Although these are important problems, they came under only tangential discussion in the seminar and hence are only treated briefly here. They are already areas of concentration for many specialists. The third general topic in this section is language change. Although this term refers to the speech community rather than the individual, it will become apparent that the processes at work have their loci in the nervous systems of many similarly constituted individuals. The treatment of each of these problems is ‘psycholinguistic’ in that relationships between the changing structures of messages and changing behavioral organizations of message users are stressed, and the underlying commonness of these problems will be apparent from the reappearance of identical principles, chiefly learning principles.

6.1 First Language Learning

This section attempts to apply learning theory to the development of language behavior. The major concern will be with modifications produced by the actions of persons in a given language-speaking culture in setting up models of verbal behavior, administering reinforcements, etc. This account does not give attention to the maturational or genetic features which presumably operate across all cultures and may influence rates of development, sequence, individual differences, and the like. This, of course, in no way implies that these are not important, but reflects, rather, limitations in seminar time and report space. Findings in the area of motor skill development mark this as an important area of research, and it is being carried on by such men as Jakobson, Leopold, Grégoire, Ohnesorg, and Cohen. In other words, we have arbitrarily limited ourselves to the learning of language decoding and encoding behavior.

6.1.1. A Psycholinguistic Analysis of Decoding and Encoding⁶³

6.1.1.1. Language decoding. In human communication decoding refers to the process whereby certain patterns of stimulation (usually auditory or visual) elicit certain representational mechanisms (ideas or meanings) via the operations of a complicated central nervous system. The basic question here is, how do certain stimulus patterns (signs) come to represent other stimulus patterns (objects), i.e., how are meanings acquired?

The first steps in the development of meaning, and hence in learning to decode the environment, are inseparable from the first steps in the development of perception. We infer that intimate ‘knowledge’ about common objects in the environment is first obtained from their proximal cues—the sensations of warm milk in the mouth and stomach, the feeling of a wooden block in the hand, the experience of being cuddled. We further infer that, since distal cues (visual, auditory) of objects can antedate their palpable presence, these cues will tend to become signs of these objects. The unique visual cues from the infant’s bottle will become signs of milk-object, the sounds of mother’s voice will become signs of her palpable presence, and so on. The general mechanism here can be seen by reference to Figure 14. Total stimulation from the object (S) elicits a complex set of reactions (R_T); in the case of the baby’s bottle, these reactions would include sucking, salivating, swallowing, and so forth. The distal stimuli () which regularly antedate or accompany total stimulation from the object will tend to evoke some reduced portion of this total behavior as a representational mediation process (r_m); in the present instance, sight of the bottle may produce anticipatory salivating and lip-pursing movements. The self-stimulation (s_m) arising from the mediating reaction is the conscious awareness of meaning and may become associated with various instrumental sequences (R_x), such as reaching forward with the hands, vocalizing, and so forth (e.g., encoding mechanisms).

Distal cues (perceptual signs) bear a necessary and inevitable physical relation to the objects they represent—not the arbitrary, assigned significance characteristic of most linguistic signs. Since the distal cues of common objects appear in a variety of contexts—at various angles of regard, under various illuminations, at varying distances, and so on—but antedating the same behavioral object, these modes of appearance become a class of signs having the same significance. This is the phenomenon of perceptual constancy, and it is only one instance of the intimate relation between perceptual and meaningful processes (cf., section 3.1.1).

In learning the significance of linguistically coded stimuli, the representational processes already established in the child’s pre-verbal, perceptual experience with common objects and situations are merely transferred (conditioned) to those auditory stimulus patterns which adults arbitrarily assign to these objects. The typical procedure is for the adult to deliberately or unconsciously direct the child’s attention’ (orientation of exteroceptive receptors) to some object while repeating the vocal sequence which for him labels the object. It is characteristic of language that the same noise is usually applied to a class of objects and situations. The large, green, light-weight beach ball, the small, red, dense, rubber ball, the small, white, hard, golf ball and so on are all stimulus situations labelled /bohl/ at haphazard intervals. It can be shown that a hierarchy of representational mediation processes will emerge from such experiences. The strongest and most available decoding habits in the hierarchy of the auditory stimulus /bohl/ will be those most frequently elicited by the distal cues of the particular objects encountered. Since most ball-objects are round, graspable, resilient, and throwable, representations of these common characteristics will gradually become the stable significance. It should be pointed out that we are dealing here with concept formation.

Figure 14

6.1.1.2. Language encoding. In human communication encoding is the process whereby a speaker’s intentions become coded into those vocal reactions which produce intelligible sounds in a given language. This is commonly called the ‘expression of ideas.’ It involves both the formation of complex motor skills and their association with representational mechanisms of the sort discussed above.

(1) Development of vocal skills. The development of basic vocal skill components in young infants can probably be viewed as a gradually changing series of stages. The first stage we might typify as ‘random’ behavior. We know for other easily observed and recorded motor systems that the earliest activity is simply a kind of mass activity. The system does everything it is capable of doing, as if the motor neurons were firing off indiscriminately. This appears to be true of verbal behavior as well. Profiles of sounds produced by new-born infants show, no differences over racial, cultural, or language groups. The determiners of frequency of emission of given sounds appear to be physiological rather than situational. When by happenstance the articulators are in a given position, a given sound emerges. As the organism develops, a progressive differentiation seems to take place (again, as with other motor systems). More and more aspects of the verbal production become ‘differentiated out’ and controllable as indicated by repetitions, predictable variations and the like. It seems probable that the gross features of the behavior are controlled first. Thus, we should expect volume and pitch control much earlier than precise articulations.

As further development occurs, control and the possibility of repetition and persistence extends to the fine musculature of the articulators, and at about this point we begin to talk about the ‘babbling stage.’ Analysis of sound profiles here indicates that differences are evident between infants in different language groups. How do these arise? One of the phenomena of learning discussed earlier was that of secondary reinforcement. Stimuli which have been present during or preceding reinforcement acquire reinforcing power themselves. We may assume that parents’ activities and, indeed, the parents’ presence, is reinforcing to an infant, and that this reinforcement has been accompanied by verbalization on the part of the parents in the sounds of their own language. Thus, the language sounds themselves can acquire secondary reinforcing power, e.g., become signs of primary satisfactions. We may predict, then, that when a child utters a sound like one in the language, this act through its auditory feedback is in itself reinforcing. This response is increased in strength over the utterance of sounds which do not appear in the language and hence have less reinforcing power. When we say that the strength of an utterance is increased, we mean that its probability of appearance is raised, and we would predict that a given sound (at this stage) will be made over and over again until it is temporarily extinguished—at which point another rewarding sound will take its place.

It has long been postulated that a ‘circular reflex’ of some sort accounts for babbling behavior (cf. E. B. Holt or F. Allport), but it is questionable whether this is needed to explain babbling, and it is obviously not sufficient to explain the change in sound patterns toward those of the culture. The writers incline to the notion that secondary reinforcement is a necessary and sufficient condition to explain this phenomenon.

Accepting the fact that gratifying speech sounds tend to be selected from a larger potential pool of skills, how are these particular spatiotemporal integrations established and strengthened? Motor synchronizations and sequences that occur on reflexive, echolallic, or imitative bases with sufficient frequency are a sufficient condition for establishment of proprioceptive, feedback systems. This makes possible more rapid and stable occurrence, which is the necessary condition for organization of these skill components on a central motor level (cf., section 3 of this report). In any case, it seems doubtful if phonemes are ever produced as isolated units, except by a process of abstraction—babbling (the earliest stage of ‘deliberate,’ repeatable encoding of speech sounds) is typically syllabic rather than phonemic.

The language community exercises continuous control over the variation which is permitted for any speaker. A person’s language is susceptible to external pressures as well as to the internal pressures toward modification. If a difference makes a difference (i.e., if it is phonemic), there is a concerted, (though perhaps indirect), effort on the part of the community to enforce a discrimination—people either do not understand or misunderstand. The language community is thus a source of direct differential reinforcement. The learner himself has a source of control in that he gets continuous feedback from his own productions and can compare those productions with those of others in the community. Diverse general social pressures may motivate him to conform, or his past rewards for conformance to models may hold him within certain distortion limits.

It seems likely that variation will be countenanced by the community where the difference does not make a difference in the code. Allophonic variation is permitted since the environmental probabilities make clear what is being said. In a similar manner the community may tolerate considerable degrees of distortion from non-native speakers (foreign accents) when the variation is consistent, predictable and, in effect, translatable. It is interesting to speculate on the effect of variations which might be phonemic in another language but not in the one being spoken at the time (say, the intrusion of a click into English). It seems likely that such variations would simply have the effect of static or ‘noise in the system’ and, depending on the degree of disruption of communication, might be tolerated.

(2) Development of semantic encoding. While it is true that certain automatisms (reading aloud, reciting number series, and the like) may short-circuit the interpretive process, most communicative acts, including voluntary speech, are largely determined by what we may call the ‘intentions’ of the speaker, for lack of a better term. We identify ‘intentions’ with the self-stimulations produced by representational mediators. The problem here will be to describe the ways in which representational mediators (meanings, significances, ways of perceiving) become associated with vocalic skill sequences.

We have already noted that the syllabic babbling responses of the infant produce auditory feedback or self-stimulation. If the same response is immediately repeated, as is the case in babbling, this auditory input will tend to become associated with elicitation of the response. This stage in the development of semantic encoding is shown as (a) in Figure 15. S_x refers to the indeterminate conditions which initiate the first occurrence of the vocal skill component (R_v), producing auditory self-stimulation . The importance of this stage is that it readies the child for imitation of ‘other person’ speech. Once vocalic self-stimulation is associated with production of the vocal reaction, auditory stimuli received from other people , particularly if they are similar in stimulus quality, will also elicit the reaction, as show in Figure 15(b). This is a case of primary stimulus generalization. If these imitative reactions are rewarded (as they definitely are in the average family situation), the child develops a broad tendency to imitate other people’s speech (cf., Miller and Dollard, Social Learning and Imitation, 1941, for details here). Given this tendency to imitate vocalic responses, the remainder of the process follows quite simply. As shown symbolically in Figure 15(c), the adult vocalizes the correct label for a common object or situation while directing the child’s exteroceptors (usually visual) to the appropriate stimuli. The child imitates this sound (R_v). If, as is likely at this stage in development, the distal stimuli of the object have already acquired meaning (cf., earlier section on decoding), the self-stimulation from this mediation process (r_m ——— s_m) must also be becoming associated with the correct vocalic reaction (step 4 in the diagram). This is a unit in semantic encoding, a learned association between an ideational representing process and a particular vocalic skill sequence.

Figure 15

An example is probably in order. Through his babbling practice, the child will imitatively produce /boh/ upon hearing his mother say /bohl/—note that perfect imitation is not expected, since accuracy depends on the babbling practice the child has had. On repeated occasions mothers and other linguistically sophisticated individuals hold up, point to, hand over, and handle the object BALL while vocalizing the label. Since the distal cues of this object already elicit representational mediators (e.g., have significance) in the child—he has already played with such objects a great deal—these intervening processes tend to become associated with his own imitative vocalization of /boh/. Note how the development of such semantic encoding frees the individual from dependence upon immediate external cues—any antecedent condition, ideational and motivational as well as external, which gives rise to the appropriate mediation process becomes capable of eliciting this bit of vocal expression (e.g., desire for the object). Being associated with a common mediation process, this vocal skill will also transfer to all those signs which elicit the mediator, i.e., the label ‘spreads’ to all members of the class, ‘ball.’

6.1.1.3. Grammatical aspects of decoding and encoding. So far we have dealt with pure semantic decoding (the association of representational mediators as responses with more or less isolated signs in various modes of presentation) and pure semantic encoding (the association of representational mediators as stimuli with more or less isolated vocal skill sequences). Most messages received and sent by adult communicators, however, involve complex sequences of signs, many of which have a largely operative function—the connective matrix within which semantic units are studded. This matrix of non-lexical material follows complicated but largely unconscious rules; neither speaker nor hearer is normally aware of deliberately selecting or noting word orders, appropriate affixes, and the like, yet these phenomena proceed in highly predictable fashion. This is probably part of what Sapir had in mind when he referred to ‘the unconscious patterning of language.’ These rules are the grammar of the language.

That the decoder is reacting to grammatical information, however, is indicated by the sharp awareness to error when wrong signals are received. The absence of an s-ending on the verbs in the boy who live next door eat the apple delivers an error signal to the sophisticated listener. Some process—set in motion by the singular noun form, persisting through reception of the verb form, and predictive of the nature of this verb form—must be postulated to account for this sensitivity to grammatical error. Similar postulation must be made to account for the encoder’s unconscious precision in ordering and affixing tags through long conversational sequences.

Grammatical facility is here assumed to be a special case of the formation of anticipational (decoding) and dispositional (encoding) mechanisms in the human nervous system, both dependent upon the frequent repetition of redundant events in sequential inputs or outputs. Following Hebb’s general notion about neural integration,⁶⁴ according to which near-synchronous activities in neighboring loci tend to become more and more strongly associated by the development of synaptic connections, if two or more input or output events, a and 6, are both redundant (e.g., occur together or in close sequence) and frequently experienced, the central neural representation of one will tend to become a condition for the occurrence of the other. Under conditions of very high frequency and redundancy, a may become a sufficient and hence ‘evocative’ condition for the occurrence of b (and certain perceptual errors and illusions may result, for example); under lesser degrees of frequency and redundancy, a merely becomes ‘predictive’ of 6, by lowering the threshold for the occurrence of b in competition with other possible events (and increased stability of decoding or encoding sequences is thereby provided).

(1) Ordering grammatical mechanisms. At an intermediary stage in language development the child encodes largely in holophrastic units (e.g., ‘pure semantic encoding’). Each expression is a content unit. The ‘little words’ (connectives, prepositions, articles) are lacking as is grammar in general. Similarly, we find that certain aphasic individuals encode in what has been called a ‘telegraphic style,’ again lacking connective words and grammatical correctness. The normal adult both orders his semantic units into certain arbitrary constructions and surrounds them with grammatical ‘cement.’ In the argot of electronic computers, we might say that global impulses from the semantic unit feed into a ‘sequence timer’ which unreels the message in a certain order and adds certain elements according to the rules of its construction.

Let us take a series of utterances all fitting the same standard syntactical construction: (the fat man) (is walking) (on the sidewalk); (a black dog) (was running) (after a car); (awful storms) (come) (in the fall). Every time the child uses such a construction, perhaps in deliberate imitation of adult speech, definite sequences of events take place—adjectival forms are followed by nominal forms, these in turn being followed by verb forms, which are themselves followed by prepositional phrases. We have here the general condition for formation of an ordering type of disposition, whereby activity type a in the motor associational area ‘tunes up’ activity type b, and so on. Having said the big, fat. …, we experience a disposition toward encoding some nominal form; having encoded an actor- action phrase, e.g., boys eat..., we experience a hierarchy of multiple readinesses for several subsequent types of phrases—prepositional (in the woods, with their hands, etc.), object (food, meat, etc.), adverbial (rapidly, quietly, etc.), and so on. The comparative strengths of dispositions in such hierarchies could be estimated empirically by presenting incomplete utterances of the type given above. Similar mechanisms operate in sequential decoding.

(2) Set mechanisms. In lengthy sequences of encoding the speaker may have to maintain and reiterate, for the hearer’s benefit, certain types of information. An entire discourse may be cast in some past time, may be concerned with plurality of number, deal with feminine gender, and require a subjunctive mood. It would obviously be convenient for both speaker and hearer to delegate these constancies to some lower-level, automatic mechanism. In computer language, the mechanism here would be analogous to a kind of ‘locking device’ which sets the computer for some repetitive operation, say, to add two zeros to any number ending in 5. Similarly, when the ‘past’ oriented human communicator encodes any verb, a dispositional set operates to add some one of the allomorphs of -d.

Set dispositions also participate in hierarchial habit systems. Whenever a variety of semantically determined contents (stems or roots of words) channel upon the same dispositionally determined suffix, we have a convergent encoding hierarchy. Examples would be plurals (lamp/s, root/s, cat/s, stem/s, leave/s, boy/s), verb tags (go/ing, sing/ing, play/ing), adverbial tags (smart/ly/, casual/ ly /, soft/ly), and so forth. Such convergent hierarchies are the psychological condition for generalization (transfer, spread of habit); we can immediately see a feasible mechanism here for what linguists refer to as extension by analogy. Having established boy: boys, cat: cats and so forth for the common signs in early life, this paradigm generalizes promptly to argot: argots and system: systems as these new nominal contents are encountered. The same generalization mechanism applies to syntactical constructions; a construction formed with such simple contents as the doggie eats his dinner quickly generalizes to such complicated contents as the linguist studies his corpus assiduously.

Wherever the same semantically determined content terminates in varied dispositionally determined suffixes, we have a divergent encoding hierarchy. All of the suffixes which can be combined with a given root morpheme constitute such a hierarchy, e.g., play/-, play/s, play/er, play /ing, play/ed, play /fulness, play /fully. Since divergent hierarchies are known psychologically to produce interference, the question arises as to why interference is largely lacking here. Encoders rarely substitute erroneous grammatical endings, saying the child player in the field for the child plays in the field, for example. This highlights the discriminatory function of dispositional sets in encoding. Interference occurs within hierarchies only to the extent that highly similar stimulus situations are operative. Note that in each case where a different suffix is applied above, a different dispositional set is operative. In other words, in each case the same semantic stimulus input is compounded with a distinctive dispositional input, making it possible to encode different endings discriminatively.

It is clear what should happen in cases where both the semantically determined content and the dispositionally determined set are the same (or highly similar), but a divergent set of suffixes must be encoded. This is the condition for interference and hence errors. The prime examples of this occur with the irregular forms of a language. With a constant dispositional set (plurality) and semantic determinants similar to other regular forms (leg: legs; toe: toes; hand: hands), the youngster typically encodes foots as the plural of foot—and similarly for goose: gooses; mouse: mouses, and the like. With a constant dispositional set (past tense) and semantic determinants similar to regular forms (walk: walked; play: played), the youngster typically encodes breaked as the past of break and catched as the past of catch. An interesting prediction arises here: since the shift in such divergent hierarchies is typically from a weaker to a stronger habit, one can predict that errors in encoding irregular verbs will be inversely related to their frequency of usage in the language. Since go is a very high frequency verb, there should be less tendency to encode goed as the past form than to encode slayed (rather than slew) as the past of slay, a relatively infrequent verb. In other words, only relatively high frequency words should be capable of persisting in their irregularity against the combined onslaughts of regular dispositional tendencies.

(3) Congruence mechanisms. There are many characteristics of grammatical encoding and decoding that simply reflect regularities in the message itself. These are the various correspondences or agreements that are maintained between parts of messages—a kind of useful redundancy. Most familiar to English-speaking communicators, perhaps, are the congruences set up between nominal and verbal forms. In the present tense, a singular noun takes a verb form ending in -s while a plural noun takes a verb form with a zero ending (the boy eats; the boys eat.). We must postulate that such congruence mechanisms have a reverberatory ‘holding’ characteristic; set in motion by the occurrence of a prior grammatical tag, they persist ‘silently’ until a second, corresponding tag is encountered, whereupon they release a particular encoding unit and are themselves eliminated. In computer language, this operates much like a condenser, one input signal setting up a cyclical action which is only released into another channel when another input signal of a specified type is received.

One of the characteristics of reverberatory mechanisms in the nervous system is that they tend to extinguish in fairly short order unless reinforced repeatedly from some external source. This leads to the prediction that, other things equal, the longer the delay between congruent elements in a message the greater the probability of error. Whereas the probability of encoding the boys runs fast is relatively low, the probability of the boys whose father used to be a track star runs fast is greater. Confusion among the prior tags for a congruence relation also produces errors, e.g., singular and plural forms occurring close together and prior to a verb as also illustrated in the sentence above.

6.1.1.4. Summary. A model for both decoding and encoding operations in human communicating has been developed on the basis of a mediational type of learning theory. On the decoding side, linguistic messages are viewed as sequences of auditory (or visual) stimuli which include cues for both semantic and grammatical decoding operations; on the encoding side, similarly, the vocalic (or orthographic) skill sequences are jointly elicited by stimulus compounds arising in both semantic and grammatical levels of organization. The semantic or ideational level of organization has been identified with the development of representational mediation processes’, the grammatical level of organization has been identified with the development of anticipational (decoding) and dispositional (encoding) processes; the motor skill level has been identified with integrations of sequential activities in the motor projection area itself; and the message level comprises either the patterns of auditory and visual stimuli received by the communicator (decoding) or the patterns of vocalic and orthographic skill sequences produced by the communicator (encoding).

6.1.2. An Experimental Analogue for Studying Language Learning⁶⁵

As a means of studying problems of language acquisition, the seminar tried to devise a simplified model of language behavior, such that experimental manipulation and control would be feasible. The experimental model should include the essential characteristics of natural languages, and the following are at least minimal requirements: (1) The model must have both a complete dictionary and a complete grammar. (2) It must be constituted of a hierarchy of units, such that every ‘utterance’ is completely organized in terms of any of the levels of the hierarchy. (3) There must be units which are sequential and units which are synchronous in respect to each other. (4) Operation of the model must involve both perceptual, receptive processes and activational, responding processes. Corresponding to the aural-oral cycle of speech in natural languages, our model will be based on a visual-manual cycle.

If we want to use our model to study acquisition of first languages, we must eliminate, or at least carefully control, the possibility of mediation by translation. Three suggestions were made as to how this might be done: (1) Use of chimpanzees, or lower primates as subjects. This might be too expensive, too slow for complex problems, and possibly not parallel to human language learning (although it might be of interest in its own right). (2) Human infants might be used as subjects, but their general nonavailability for prolonged and rigidly controlled experimentation makes this possibility infeasible. (3) The solution actually adopted is to put adult humans into a non-translation situation, that is, where mediation by their own language would be not necessarily absent, but constant and trivial. Of course, with adult subjects the problem situation cannot be identical with the situation of infants learning their first language, at least from the point of view of the subjects, but we can control what we believe in theory to be the essential factors of the situation, and the model can easily be changed to follow any necessary modifications in the theory.

The subject (who may suggestively be called ‘infant’) sits before a panel below which is a set of control knobs or levers. As he sits there, lights of different colors and in various sequences flash on the panel in front of him. The flashing of lights is controlled by an experimenter (‘parent’) sitting at a corresponding panel in another room, unseen by the infant. The parent’s ‘messages’ are ‘meaningful,’ i.e., are in accord with a system of intent as coded in the model language, but it is probably not necessary that the infant be preinformed that the light flashes he is watching are so meaningful. He may respond to the flashes in a variety of ways, but eventually he will probably try moving some of the controls and see the visual result of his own responses (e.g., ‘feedback’). This would correspond roughly to the random behavior stage in vocalizing. Some mechanism of reinforcement (say flashing the word ‘right,’ which we assume has reinforcing properties for the subject) strengthens such responses. While such reinforcement may be used initially just to get the subject to ‘say’ something, later it serves in the discrimination between proper and improper responses for the code. This parallels the increasingly discriminated reinforcement applied in language learning.

The infant watches the patterns of light on the panel, and with reinforcement presumably will attempt to duplicate these, or make any other reinforcible set of responses, by manipulating the controls at his disposal. Patterns of increasing complexity will be discriminated by the infant, and hierarchy of patterns according to complexity will be established both for his recognition and for his motor manipulation. Eventually, we would suppose that he will learn the entire code, the model language.

The apparatus consists of two light-boards with controls such that operating controls on one board turn on lights on the other board, and if we want visual feedback, lights on the same board as well. The effects of getting feedback which differs from the parent’s signals can be studied by building in a different set of relays of the controls to the two boards. The controls are of two types, corresponding approximately to natural language segmental and supra-segmental features: the left hand operates controls which are held constant over more or less larger sequences; the right hand may operate one of two other types—either controls (perhaps press bars) which are operated in sequences, corresponding to segmental phonemes, or a single complex mechanism with independently but synchronously acting elements, corresponding to distinctive phonemic features, e.g., a knob which may be pushed or pulled, raised or lowered, moved right or left, or twisted left or right, with neutral positions possible for any of these operations. Operating these controls produces patterns of light on the boards. Visual qualitative differences between infant’s and parent’s patterns may be introduced by such modifications as intensity differences or wave-length shifts which contrast the lights produced by parent and those produced by infant. These differences parallel the non-linguistic features which enable us to distinguish one speaker’s voice from another or other’s voices from our own.

Since the model is to be simpler than natural language in nonessential points, the code may be constructed with a simpler ‘phonology’ (three or four distinctive features, or perhaps five phonemes, plus two suprasegmental features), morphology (perhaps 30 morphemes in three classes as determined by phonemic constituency or position of occurrence in the sequence), and syntax (two possible orders of morphemes). Of course, the code can be complicated to any desired degree, and the corresponding effect on learning can be measured.

The problem of how subjects can be motivated in the rather laborious set-up proposed here was raised. It was generally agreed that we cannot tell until the experiments have been tried whether the satisfaction of ‘solving the problem’ would be adequate in maintaining the cooperation of the subjects. It may be true that children in the real situation are subject to the same boredom as the ‘infants’ in the experimental situation, so this factor might well be studied by experimentally varying the type of reward given the subject for participating in the experiments. In general, however, it appears likely that some sort of relevant reinforcement must be amply provided.

It will be noted that no ‘meaning’ is attached to the code symbols or patterns of symbols, as they appear in the experiments outlined above, other than that they belong to a code, are recurrent, and are somehow differentially reinforced. If ‘referential’ meaning is desired in the system, it can be introduced in a number of different ways. One way suggested was to introduce a separate field on the light-board—a field containing a moving light. Certain patterns of light in the vari-colored message-field, could be associated with movements of the light in the ‘semantic’ field, ‘requests’for action on the moving light, etc. Obviously, any simple modifiable visual or auditory stimulus could be used as the referent. Since in natural language, many complicated conversations are carried on by response to linguistic and not referential material completely (note: Hello. How are you? Fine, thanks, etc.), it may be that the mechanism of language can be studied for our purposes without the many complications introduced by reference to extra- linguistic things. Furthermore, as we have previously noted, we may complicate and interfere with the learning process by the mediation of translation.

Although designed primarily for study of problems of first language learning, the apparatus can be used for study of different sorts of problems. For instance, having thoroughly mastered one code, the infant can be given proper instructions and run through the experiments again with a different code. In some ways this situation can be made to resemble second-language learning situations, but it is doubtful that this experimental design is the best suited for study of such problems. It has also been suggested that interesting data might result if two subjects equally ignorant of any prepared code were placed at the two panels, rather than a naive ‘infant’ and a trained ‘parent.’The nature of the communicative code, if any, finally adopted by these subjects might shed some light on their conception of their own natural languages.

Some specific problems in language learning that might be studied in this situation are discussed below. One problem previously referred to is the role of secondary and generalized reinforcement in language learning, particularly their role in the approximating of cultural sound patterns which seem to occur with the babbling stage. In the experimental situation outlined above we might ask what the likelihood is of a subject repeating a ‘model’ pattern of lights under varying conditions. (1) Repetition when the patterns appear along with nothing else in the stimulus field. This would presumably give a summary measure of the subject’s past reinforcement history with respect to imitation in this kind of situation. (2) Repetition when the patterns appear at the same time that reward is given. This would give the patterns the status of secondary reinforcers. (3) Repetition when the patterns appear initially without other stimuli but with reproductions by the subject being reinforced. We would here have a measure of the rate at which the subject learns to imitate initially neutral models. At this stage we would perhaps be most interested in (2), the reproduction of secondary reinforcers. This case seems to be closely parallel to the postulation earlier concerning babbling. While, of course, work with adults (even if successful) would not deny the existence of the circular reflex in infants, it would clarify the possibility of the alternative explanation and might lead to further, more conclusive research.

A second group of problems which might be attacked here would be those related to certain transitional phenomena. For example, it can be hypothesized that set dispositions, ordering dispositions and congruence dispositions (tense, order and agreement) are all dependent on frequencies with which sequences appear in the code and manifest themselves in the same way in association tests, recall situations, etc. Our experimental situation would permit us to study the growth of such dispositions and to control the actual frequencies in any manner we choose. Experimental analogues of association tests and recall tests can be easily constructed.

Further consideration along these lines would lead us to another set of experiments on paradigmatic development (see the section 5.4 for a discussion of the influences of paradigmatic and syntagmatic relationships upon association). One very basic question which might be asked is whether we could build up ‘use by analogy’ in our experimental model. Suppose in one sequence of learning trials we presented units singly at first and then in later stages began using the units in context. Suppose unit one were rewarded in context one and unit two also rewarded in context one. A new context might be introduced and again both units rewarded when they appear in the context, and so forth. Then, if unit one were presented in context ‘X,’ would the subject tend to encode unit two in that wholly new context? This learning-by-analogy is often assumed in descriptions of language processes but has rarely been attacked experimentally. If this can be demonstrated, then another very interesting question arises—will unit one evoke unit two in an association test? Here we could have a control over the paradigmatic class which is not at all possible in ‘natural language.’ On a broader scale, studies of analogy such as these might be extended to verb tenses, sentence forms, etc.

These are only a few examples of the kinds of learning studies which need to be performed. Whether this situation will approximate language in such a manner as to lead to fruitful results is, of course, a matter for empirical confirmation. One seminar member objected that this was ‘nothing more than complex instrumental learning’ but that is probably exactly the way an arch-behaviorist might describe language itself. In general, we might say that this is a highly flexible coding and learning situation which provides for ‘listener,’ ‘speaker,’ and ‘listener-speaker’ situations. It has advantages over the experimental use of an artificial language in ease of production of the stimuli (it doesn’t need a trained speaker or phonetician), ease of control of the code, and ease of precise recording (each response is clearly what it is and can be set to make its own record).

6.2 Second Language Learning and Bilingualism⁶⁶

When, after becoming a practical expert in his own, first language, a person starts learning a second language, new sets of decoding and encoding habits are being formed in competition with the old. When the bilingual shifts from language to language, similarly, two systems of decoding and encoding habits come into conflict to a greater or lesser degree. The fact that the same general principles found to be important elsewhere in this section also are significant here justified discussing second language learning and bilingualism in the present context. Since the seminar did not devote much time to these topics, however, only a brief sketch of the thinking of some of us on these problems is offered. The reader is referred to recent books by Uriel Weinreich and Einar Haugen⁶⁷ for excellent treatments, undertaken from the linguistic point of view, but with very considerable psychological and sociological sophistication.

6.2.1. Compound and Coordinate Language Systems

Both second language learning and bilingualism involve the acquisition and utilization of two linguistic codes. The messages produced in the two or more languages employ differently constructed and organized units, different grammatical rules, and different and equally arbitrary lexical systems, excepting occasional cognates. To the extent that phonemic systems are different, two sets of differentiations and constancies on the decoding side and two sets of vocalic skill components on the encoding side have to be maintained. Since the entire systems of transitional redundancies in two languages are different, alternative anticipational and dispositional integrations have to be established. And since the lexical aspects of messages in two languages are different, alternative sets of semantic decoding and encoding habits have to be maintained—in other words, alternative sets of associations between message events and events in the representational system, or meanings.

Perhaps because of dependence on the model provided by second language learning in school situations, many writers seem to have assumed that meanings are constant in second language learning and in bilingualism. The meaning of the object HORSE remains the same as perceptually experienced. Hence the meaning of its alternative linguistic signs, horse/Pferd, must be the same—all that is involved is two systems of coding the same meaning. This is the case under certain circumstances, which establish what we shall call a compound language system. In such a system, as shown in Figure 16, two sets of linguistic signs, one appropriate to language A (_A) and the other appropriate to language B (_B) come to be associated with the same set of representational mediation processes or meanings (r_m → s_m). On the encoding side, likewise, the same set of representational processes comes to be alternatively associated with two sets of linguistic responses, one in language A (_A) and the other in language B (_B). This development is typical of learning a foreign language in the school situation. It is obviously fostered by learning vocabulary lists, which associate a sign from language B with a sign and its meaning in language A. A compound system can, however, also be characteristic of bilingualism acquired by a child who grows up in a home where two languages are spoken more or less interchangeably by the same people and in the same situations. In this instance some compromise representational processes taken from both languages may be established, with neither having pronounced dominance.

Figure 16

A very different kind of relation between two languages in the same nervous system is what we shall call a coordinate language system. In this case, as shown on the right-hand side of the diagram, the set of linguistic signs and responses appropriate to one language come to be associated with one set of representational mediating processes (r_m1 → s_m1), but the set of linguistic signs and responses appropriate to the other language become associated with a somewhat different set of representational processes (r_m2 → s_m2). This kind of development is typical of the ‘true’ bilingual, who has learned to speak one language with his parents, for example, and the other language in school and at work. The total situations, both external and emotional, and the total behaviors occurring when one language is being used will differ from those occurring with the other. The kinds of representational processes developed must then also be different and hence the meanings of the signs. This development can also characterize the second language learner, who, relying as little as possible on translation and immersing himself in the living culture of another language community, comes to speak a second tongue well.

Even within a coordinate system there may be interference between the two sets of processes. Given the likenesses throughout human cultures in the situations and objects dealt with by language, it is certain that the representational processes elicited by translation-equivalent signs in two languages wall often be similar. In decoding, this produces a constant pressure on the bilingual to confuse meanings, to interpret a sign in language A as its translation-equivalent in language B would be interpreted. The more similar the signs—cognates, for instance—and the more similar the mediators, the greater this pressure will be. Interference is most likely to occur when the languages are closely related and the cultures or the experiences associated with the languages are alike.

On the encoding side, the more similar the meanings or representational processes, the more errors there will be. These may consist in delays or blocking of response, if the alternative responses in the two languages are quite different. There may be intrusions of responses from the wrong language, if the items in the two languages are similar. These phenomena are often obvious in the compound system, where identical mediators must elicit alternative responses. They may take subtle forms in the coordinate bilingual, resulting merely in minute delays or shifts in response frequencies in comparison with monolinguals. Compromise formations usually result, depending upon relative habit strengths in the two languages in vocalic skills, lexical associations, and transitional patterning.

In spite of the pressures for interference, there are many instances of remarkably pure bilingualism, in which the speaker, once launched in a given language, in an appropriate situation, and speaking of events associated with that language, will experience no difficulties and perform like a monolingual. There are at least three general predictive factors to be considered for the coordinate bilingual. In the first place, the feedback stimuli from previous utterances in a given language are more associated with mediators appropriate to that language than another, unless considerable language mixture has occurred in past usage. Secondly, the current interpersonal situation will affect interference in speech as much as it does features of style or dialect within one language. Hearer bilingualism, the relative prestige of the languages, momentary feelings toward the hearer will alter the general availability of responses in each language or even lead to deliberate use of interference by a speaker. Finally, stimuli arising from the scenes, objects, and people present during the formation of a language will also be associated more strongly with mediators appropriate to that language. Hence bilinguals report that when they are with, or even think about, their parents or their home, the parental language becomes more available. A bilingual under emotional stress may revert to the language spoken when comparable emotions have been experienced in the past.

For any semantic area we would expect speakers of more than one language to distribute themselves along a continuum from a pure compound system to a pure coordinate system. How would one measure or index the location of particular individuals along this continuum? At one extreme, the meanings of translation-equivalent signs are identical, and at the other the meanings of translation-equivalent signs are different. Furthermore, the semantic differences involved tend to be connotative rather than denotative. The semantic differential (cf., section 7.2.2) seems particularly appropriate as a tool here. It could be used to measure coordinateness within a semantic area or to make a general estimate if appropriate samples could be devised. If a sample of pairs of translation-equivalent signs were given to a varied group of two-language speakers for differentiation against an appropriate form of the semantic instrument, with D between profiles of the pairs computed for each speaker, the average D (difference in meaning) should vary directly with the degree of ‘coordinateness’ of the language systems within each speaker. The validity of this measure could be estimated against such criteria as frequency of interference in ordinary conversation, fluency measures, and translation facility, to which we now turn our attention.

6.2.2. Translation under Compound and Coordinate Systems

In the process of translating from one language to another, linguistic signs in one language (_A) must be decoded and equivalent or related linguistic responses must be made in the other language (_B). The behavioral situations are quite different, depending on (a) whether the translator maintains compound or coordinate languages in his nervous system, and, if the former, (b) whether he is translating to or from his dominant language. The left-hand diagram in Figure 17 represents the translation process for a compound language system. Solid lines show translation from the dominant language, and dashed lines show translation to the dominant language. The right-hand diagram represents the translation process for a coordinate language system; the encircled numbers represent alternative translating circuits at different stages in the development of translating fluency in the coordinate system.

Compound system translating. When the product of an ordinary foreign language course in high school, let us say, translates from his native English into the other language, we have the situation represented by the solid lines in the left-hand diagram. Encoding the foreign forms involves direct response competition, since both are associated with a single set of American culture meanings. The ‘same’ mediated stimulus must elicit a response (foreign language output) quite different from the dominant response (English language output) in the habit hierarchy. The task would be impossible were it not for differences in the total stimulus pattern for the translator. Such differences may be brought about by the ‘set’ to translate, the feedback of foreign cues from preceding output, distinctive dispositional tendencies (once the foreign language grammar and syntax has become sufficiently learned), and unique associations with the use of the second language. These cues must be sufficient to counteract the stronger English response tendencies in the presence of s_m. An analogous situation exists for the bilingual who has learned to speak both languages in a homogeneous environment. Encodings in either language in response to inputs in either language maybe equally reinforced. One would expect more or less continuous interference and intrusion in encoding, in either language. This interference is not due to the pressure of a dominant language, as in the earlier example, but to lack of distinctive cues in previous use of the two languages.

Figure 17

Theoretically, decoding from a foreign language should be facilitative for a compound system, since different inputs are associated with the same representational reactions or meanings. To the extent that signs in the two languages are similar in form, decoding will be made easier. In a sense, the process of decoding the non-dominant language always involves translation, since the representational processes of the compound system are appropriate only to the dominant language. It is interesting to note that the compound system, whether a result of second-language learning or childhood bilingualism, can never translate in a true cross-cultural sense, since there is no possibility of comparing meanings in two culture contexts when only one system of representational processes is present.

On the other hand, the coordinate system provides true cross-cultural translation, but with certain theoretical complications, for the system itself is changed by the translation process. Let us take the ideal case of the bilingual in whom both languages are of about equal availability. As shown in the right-hand side of the diagram above, signs in language A must elicit meanings appropriate to language community A, due to the processes of learning to decode in a coordinate system. However, these language A mediation processes are associated with encoding in the same language. How then does the coordinate bilingual translate?

It will be recalled that to the extent that human cultures, situations, and objects are similar, the coordinate system is likely to involve some representational process in language B that is closely similar to that being elicited by the input in language A. This process may not be exactly the same as that elicited by the sign in language A. Translation-equivalent signs in the two languages may elicit slightly different representational processes due to the subtle differences in context or connotation which ordinary translations must lose, e.g. the difference between ‘Weltanschauung’ and ‘philosophy of life.’ Or there may be several processes in language B which are similar in different respects to a representational process in language A. Such is the case when a translator must choose between several partially adequate translations, as in the various English translations of ‘gemütlich.’ In both of these instances, simple generalization between s_m1 and s_m2 by pathway (1) may make translation possible. Since generalization is most probable between maximally similar representational processes, the result is the encoding of the most adequate cross-cultural translation.

Another possibility is that pathway (2) between s_mi and r_m2 is established by learning, which would be necessary if s_mi were so closely associated with a cultural context peculiar to language community A that appropriate generalization to language B processes was inhibited. Such learning might occur, for example, if a bilingual is in a situation which elicits s_mi. This may be due to an appropriate external context, thoughts about events that occurred in the context of language A, or a conversation with a language A monolingual. For learning to occur, r_m2 must also be elicited, perhaps because a language B monolingual describes or refers to situations or objects represented by s_mi but using language B signs. Generally, it would appear that a learned association between s_mi and r_m2 probably only develops in situations in which some generalization between s_mi and s_m2 also occurs. Learning then facilitates the appropriate translation. Whether through generalization, learning, or both, the translation is such that the meanings elicited in monolinguals in languages A and B are as similar as possible.

The ideal translator or interpreter accomplishes a transformation of signs through a three-person channel (monolingual A—coordinate translator—monolingual B) such that the representational processes of all three, or the meanings of the signs, remain unchanged. Obviously, the more the cultures, or the situations and objects discussed differ, the less rapidly the interpreter can encode. He is delayed by lack of quick generalization to similar meanings in the other language, or by conflict between several partially appropriate meanings. Interference in decoding may be produced by semi-cognates, similar or identical forms with varying meanings.

The above description treats the coordinate translator as though he were translating for the first time, but the translation process itself brings about new learning. Practice may reduce the capacity for cross-cultural translation, as we will demonstrate. Three different kinds of short-cuts may develop in the proficient translator or interpreter. In encoding, a linguistic response in language B must repeatedly occur in close sequence with the representational process appropriate to language A, as shown by pathway (3). This produces the same kind of interference discussed earlier, in the compound system, since s_mi may elicit different responses depending on whether the speaker is supposed to speak language A or to translate it. Note that the representational process of language B has not been elicited at all. The second short-cut may consist in direct decoding into language B by pathway (2). The representational process appropriate to language B occurs soon after presentation of the linguistic sign in language A when there is frequent translation. This sign in language A must therefore become associated with a mediation process appropriate to language B, and it will compete with the process appropriate to language A, and, to the extent that they are similar, will lead to elimination of the meaningful discrimination in decoding. Finally, the mediation process may be only secondary, the response in language B being directly associated with hearing of a sign in language A, without intervention of a meaning process. This is most likely to happen for simultaneous interpreters who always use the same translation for the same word or phrase.

If a coordinate bilingual is hired to interpret in one direction only, from language A to language B, we must predict that (a) he will gradually develop an appropriate set of translation meanings for the signs in language A, and (b) he will lose his ability to speak language A. In other words, he will become a perfect A-to-B translating machine, while retaining his proficiency in language B—thus, a sort of dual-input speaker of language B. If a coordinate bilingual translates frequently in both directions, he must gradually (a) lose the distinctiveness among the mediation processes appropriate to each language and (b) suffer increasing confusion in encoding. In other words, the very process of two-way translating tends to transform a coordinate system into a compound system. It would be interesting to test these predictions on a sample of professional translators over time, using the type of experimental design suggested earlier with the semantic differential. This transformation can be minimized by refreshing the monolingual associations in both languages, of course.

6.2.3. Grammatical and skill levels

We have emphasized the semantic aspects of second language learning and bilingualism because these aspects have been relatively neglected. There is a great deal of carefully analysed information on purely linguistic aspects of bilingualism, phonemic, morphemic, and grammatical interactions and the like, and the interested reader is referred to the two sources given at the beginning of this portion of the report. For the most part previous work has reported the occurrence of certain phenomena without attempting to relate frequencies of occurrence to the learning experiences of the speaker. Grammatical aspects should be particularly profitable to study from this standpoint. Being organized on largely unconscious levels on the basis of transitional redundancies, these aspects of encoding should be especially difficult to learn and, once learned, should be equally difficult to suppress when trying to master a second language. For the coordinate bilingual, on the other hand, two alternative grammatical systems once established should provide for greater stability and independence between the two language systems.

We have considered here only the aspects of bilingual speech which reflect the influence of two linguistic codes on each other, omitting consideration of such non-linguistic features as pronunciation and style, (cf. section 4.1). Yet further study of these features would probably be rewarding, especially for psychologists who would be interested in their sensitivity to differences in the attitudes of bilinguals and second-language learners toward the respective speech communities.

6.2.4. Research proposals

(1) Indexing degree of coordinateness of language systems. The possible use of the semantic differential as a means of determining the degree of separateness of meanings for translation-equivalent signs has already been discussed. (A comparison with responses of monolinguals is suggested in section 7.4.3.1. This would determine not only semantic areas in which a compound system exists, but the language community for which the meanings, are appropriate.) (2) Influence of bilingualism on perception and meaning. One of the writers⁶⁸ is now engaged in research of this nature. Subjects varying in degree of bilingualism tell stories in response to the Thematic Apperception Test (a series of rather ambiguous situational pictures), (a) in language A after preliminary instructions in that language, and (b) after an interval of several weeks, in language B after similar preparation. The expectation is that ways of perceiving these pictorial signs, their meaning to the subject, will vary with the language being used, and with the degree of coordinateness of the systems in a given subject. (3) Measuring the transitional proficiencies of second language learners and hilinguals. The ‘Cloze’ procedure developed by Wilson Taylor and described briefly in section 5 seems adaptable to problems in this area. Passages in languages A and B, as translated by maximally facile coordinate translators, could be mutilated (every fifth word deleted, for example) and given to subjects with varying degrees of bilingualism or varying amounts of second language training. In the former case, the more nearly equal the correct ‘fill-in’ scores for languages A and B, the more ‘truly’ bilingual in the coordinate sense the subject; in the latter case, the more nearly equal the scores in A and B, the greater the learning of the second language. This technique has two advantages; first, by its nature it samples all of the subtle contextual factors of both semantic associational and grammatical-dispositional levels; second, by using each subject’s own performance in his most proficient language as a criterion, it eliminates individual differences in intelligence, language abilities, and the like. (4) Measuring interference between languages under varying conditions. Three factors were cited above in accounting for the degree of interference in encoding—feedback, the interpersonal situation, and differential past experience. Encoding with these conditions varied can be studied to see how conditions influence the amount and the kind of borrowing.

6.3 Language Change⁶⁹

Languages change in response to both internal dynamics and external pressures. Psycholinguists are interested in both processes, but the present analysis concerns the former. The existence of similar forces operative under similar conditions everywhere in language is indicated by the existence of a limited number of types of change which reoccur in different, historically unconnected languages and at different chronologic periods. Linguists have accumulated an enormous amount of authenticated information relating to such changes and have been able to formulate a number of principles regarding the ‘hows’ of specific changes. The psychologist, in line with his general orientation, is typically interested in the ‘why,’ that is, in the isolation of general principles of behavior underlying shifts in linguistic habits. No doubt the interplay of factors in any particular instance is too complex to allow of complete explanation in the foreseeable future, but this does not mean that there are no general laws whose combinations limit the possibilities to the point where at least statistically probable outcomes may be hypothesized.

A striking fact about language change is that although both changes of form and those of meaning are always proceeding simultaneously, each can be extracted by analysis as a set of separate and practically independent processes. Nevertheless, they do impinge on each other at a few points. Each of these two major areas of change, the formal in which will be included both phonological and grammatical changes, and the semantic, will here be given separate treatment. In each of these two major areas, tentative generalizations concerning the facts of change based on linguistic data will be followed by a discussion of the psychological mechanisms which may be suggested as operative in bringing them about.

6.3.1. Formal change

6.3.1.1. Linguistic facts. Changes in the phonemes, the basic units of the sound systems of language, may involve (1) replacement of one phoneme by another, (2) loss of a phoneme, (3) transposition or metathesis, (4) insertion of a phoneme. These types have been stated in decreasing order of frequency of occurrence, replacement being by far the most common. Changes in phonemes may further be classified as regular or sporadic. A change is regular if it affects all instances of a given phoneme under specified conditions. Otherwise it is sporadic. Examples of regular changes are the following: In early Aramaic p in all instances was replaced by t; in (probably) 18th century Hausa, a West African language, s was replaced by s in all instances where e, e’, i, or i’ followed. An example of a sporadic change: In the development of Spanish from Latin, r was replaced by I in the word arbol ‘tree,’ earlier arbor. In cases of sporadic change, reference to specific instances is unavoidable.

Regular changes are, in turn, divided into unconditioned and conditioned. In unconditioned changes, all instances of a phoneme are affected without limiting conditions. A conditioned change involves only some of the occurrences of a phoneme under stated limitations in terms of other phonemes or positions in the utterance. In the above paragraph, the first change mentioned is unconditioned, the second conditioned. Unconditioned changes tend to occur in sets. What is involved is the replacement of some one distinctive feature common to a number of phonemes by another distinctive feature in all its occurrences. Sometimes such changes are, as it were, reversed in midstream. After affecting one or more phonemes, the older form reasserts itself in some instances and the result is somewhat checkered. An example of such a mass shift is Grimm’s First Law, a statement of consonant changes from Proto-Indo-European to Proto-Germanic. For instance, /p > f, t > J>, k > x/ among other changes, that is, in the three instances cited, a stop feature was replaced by a fricative feature.

Two types of results may be distinguished here: either a phoneme or group of phonemes through a change in feature may give rise to sounds which did not previously occur in the language, or change may be to sounds which already do exist. In the former eventuality, no change in phoneme inventory (i.e., number of phonemes) or distribution (i.e., occurrence in particular forms) results. In fact, some would call such a change phonetic, rather than phonemic. In the latter case, we have the phenomenon of merger which involves a reduction in phonemic inventory. The following empirical generalizations are offered with regard to merger. Each is understood as being preceded by ‘other things being equal’: (1) The more uncommon a phoneme is in human speech in general, the more likely it is to be merged with another phoneme. (2) The lower the frequency of a phoneme in a given language the more likely it is to merge with another phoneme, providing this second phoneme is not itself of excessively high frequency. (3) The closer the points of articulation shared by two phonemes the more likely they are to merge. (4) The more distinctive features shared by two phonemes the more likely they are to merge. (5) The fewer the pairs of different linguistic forms which are distinguished by the phonemes, the more likely they are to merge.

Conditioned changes result typically from conditioned allophonic variations in which by loss or change of the original conditioning factor a formerly nonsignificant contrast becomes phonemic. For example, in the transition from Proto- Indo-European to Sanskrit we reconstruct the following stages:

1. ka ke ki ko ku

2. ka če či ko ku (phonemically as in stage one. We now have an allophone [c] which occurs before e and i. There was no [č] in the language previously.)

3. ka ča či ka ku (a merger of o and e with a has produced a phonemic contrast ka vs. ca which did not exist before).

Another type of result ensues if a phoneme undergoes a conditioned change but the resultant of the change already exists. In this case, there is partial merger. The old instances of the phoneme which were not affected continue but the new variant merges with another previously existing phoneme. It is clear that this type of change affects the distribution but not the inventory. In recent Russian o changed to a under conditions of lack of stress. Since stressed o continued and a already existed, no new phoneme was added to the language. In general, conditioned change is the diachronic aspect of the synchronic problem of conditioned allophonic variation.

The following general facts about regular conditioned changes must be considered: (1) The conditioning factor is more often a phoneme which follows rather than one which precedes. (2) The conditioning factor is almost always immediately following or preceding. Sometimes a vowel is affected by that of a following or preceding syllable. More remote conditioning factors are very rare. (3) The change usually results in an articulation which is more like the conditioning phoneme. That is, it is assimilative rather than dissimilative. In general, then, changes result in a sequence of articulation which abbreviates or eliminates movements. For example, the fronting of a k to č before i eliminates the forward movement from back to mid position. (4) Final positions of syllables, words and utterances are often conditioning factors for change, initial positions only rarely. Changes with final position as their conditioning factor are typically those which result in loss or merger. This, of course, results in fewer phonemes in these positions. We know of few languages which have more distinct phonemes in syllable final than in syllable initial position.

Sporadic changes have the following characteristics: (1) Certain sounds are most frequently affected—liquids (r, 1), nasals (n, m) and sibilants (s, s). (2) Dissimilation is probably as frequent as assimilation, contrasting with its rarity in regular changes. (3) The conditioning factor often operates at a distance. Example: Latin peregrinus > Italian pellegrino ‘pilgrim.’ Here the succession of two r^ys has resulted in the dissimilation of the first to I even at a distance. The connection of such changes with speech-lapses, championed by Sturtevant, is highly plausible.

As has been stated, regular sound changes proceed in general without regard to the meanings of the forms in which they occur. However, it seems likely that regular changes are encouraged or inhibited depending on their consequences for grammatical structure. Thus, although both Germanic and Romance languages have stress, the unstressed vowels have tended to merge in Germanic languages but to remain distinct in Romance. This may well be connected with the fact that unstressed syllables in Germanic are typically limited to non-root morphemes, while in Romance languages the root may sometimes be unstressed, e.g., Italian ‘amo ‘I love’ but a’mo ‘he loved.’ In turn, the merger and loss of final vowels in Germanic languages has had repercussion on the grammatical system, in that distinctions based on difference of inflectional morphemes which became homonymous had now to be expressed by other, syntactic, means.

The chief process of morphological change is analogy. (1) By alternations between morphs of the same morpheme one pattern is replaced by another, usually more common; or (2) the alternation is effaced completely by extension of one of the forms to the alternant. Such changes are often stated in the form of an analogical proportion. Examples of both types of change are (1) the replacement of ‘brought’ by ‘brang’ (sing: sang = bring: brang) and (2) ‘calfs’ for ‘calves’ (cliff: cliffs = calf: calfs). The extension of a formative, typically derivational affix, to a combination in which it had not previously existed, or the formation of a new compound out of existing elements may likewise be viewed as a kind of analogical process. Thus the new form draftee can be expressed as the consequence of an analogical extension (employ: employee = draft: draftee). The less frequent types of morphological change such as folk-etymology and blending are not considered here specifically. They may be looked upon as partial analogical processes.

6.3.1.2. Certain general hypotheses relating to change in form. For convenience in discussion, principles relating to formal language change may be taken up in two categories: (1) those relating to the locus (in utterances) of change; (2) those relating to the process of change.

(1) Principles Relating to the Locus of Change. The evidence for generality described above makes it clear that changes in language structure are not haphazard as to locus—rto the contrary, there are definite ‘stress points’ within messages as sequentially unreeled. Are there any general principles, i.e., general cross-linguistically, which would make it possible to specify ‘stress points’ and hence predict the most probable locus of changes?

I. Short-circuiting. Features of subsequent phones will tend to be anticipated wherever possible. The limits suggested by ‘wherever possible’ are at least the following: (a) To the extent that an articulatory feature of a subsequent phone is incompatible in the motor sense with that of an antecedent phone in its immediate environment, the tendency toward short-circuiting will be reduced; (b) To the extent that an articulatory feature of a subsequent phone would change the phonemic (i.e., code) character of an antecedent phone in its immediate environment, the tendency toward short-circuiting will be reduced. In this case, speaker ‘lapses’ (see below) will tend to be corrected by hearers. The basis for short-circuiting within the rapidly executed and tightly bound phone sequences that constitute syllabic units probably lies in the nature of skill formation—the central programming of neural events in the motor cortex is much more rapid than the sequential execution of movements, resulting in overlapping excitation patterns and a tendency to anticipate.

II. Perseveration. Features of antecedent phones will tend to persist wherever possible. Again, ‘wherever possible’ is limited at least by (a) articulatory incompatibility, and (b) production of significant (e.g., to hearers) phonemic changes. In the latter connection, it should be noted that shifts from one distinctive feature to another are not necessarily significant—there may not exist in the language any meaningful unit (e.g., word) corresponding to the modified signal even though it is phonemic. The basis for perseveration of this kind would seem to be something akin to motor inertia or ‘least effort’; it is simply easier for the rapidly operating encoder to persist or remain in whatever articulatory feature happens to be held than to change it. Due to such perseveration, there should be a general tendency, other things equal, for phonemes displaying a certain feature to be followed by phonemes displaying the same feature in a given language. Obviously, this tendency cannot be carried to its logical conclusion, or the code becomes uniform and hence meaningless to the hearer. Elsewhere in this report (section 5.3), Saporta describes some computations on sequential phonemes in English which bear on this matter.

The combination of short-circuiting and perseverative principles in skill execution leads to the following expectations: (a) That the positions of instability in a language with reference to these factors should be those in which the phones both antecedent and subsequent to a given phone include one or more features in common. On this basis, for example, one would predict that voiceless consonants would be more common in initial and terminal positions in words than in medial positions where they would often be surrounded by voiced vowels, (b) That where a given phone is bounded by antecedent and subsequent phones displaying different features within the same dimension (e.g., voicing, tongue position, lip position, etc.), this phone should be characterized by transient or intermediary features between those in its environment. The general fact that subsequent phones have more effect upon sound changes than antecedent phones implies that the short-circuiting mechanism is stronger quantitatively, but this does not derive directly from any theoretical notion. The specification of conditions operative in determining which factor will be dominant remains extremely obscure. There is no doubt a tendency for intervocalic unvoiced consonants to become voiced—we can at least predict that this is more likely to happen than, for example, that they will become glottalized. Still, many languages go through very long periods during which intervocalic voiceless stops remain entirely stable and we are as yet unable to specify what factors, linguistic, cultural or otherwise, determine when the change will take place and when not.

A number of principles derive from analysis of transfer and interference in sequential materials. A great deal of experimental data⁷⁰ justify the following summary statements.

III. Convergent hierarchy. When a variety of antecedent states (e.g., stimuli) converge upon a common subsequent state (e.g., response), transfer is positive and retroactive effects are facilitative, the degree of facilitation varying directly with the similarity among the antecedent states. This sets the general condition for extension by analogy at the grammatical level, e.g., having learned play-ed, walk-ed, and fix-ed it becomes easier to transfer to crawl-ed, digest-ed, and master-ed, as well as for what might be called error by analogy at any level, e.g., to the extent that a given subsequent acquires high frequency with certain antecedents it should tend to generalize or transfer to other antecedents. The psychological fact that the degree of facilitation varies with similarity among the antecedent states should also have its evidence in language behavior. Certainly it is easier to extend by analogy, from boyhood, manhood, and priesthood to coinages like babyhood, warriorhood, and scouthood (semantically similar antecedents) than it would be to stonehood, lighthood, or hillhood (semantically dissimilar antecedents). On the phonemic level, until appropriate data have been analysed, one can only hypothesize that high frequency subsequents should tend to generalize among similar antecedents, e.g., there should be a tendency for the initial phonemes in sets with common terminus to be separated by fewer distinctive features than chance would dictate (time, dime and lime, rhyme, for example, are the only sets in English having this particular medial and terminal, and t/d and 1/r are separated by few distinctive features). In this connection, it would be useful to have a frequency list of syllables in English as well as other languages.

IV. Divergent hierarchy. When a common antecedent state diverges upon a variety of subsequent states, transfer is negative and retroactive effects are interfering. This is the general psychological condition for competition among responses and errors. Having learned to make one reaction to a stimulus, it becomes more difficult to substitute another reaction to the same stimulus. Observe the following:

Even for the simple task of rapid repetition, initial (convergent), medial (divergent-to-convergent), and terminal (divergent) sets become increasingly difficult in that order. On the same ground, one would expect ‘stress points’ to be located in positions of divergence rather than convergence, both at phonemic (e.g., /t/ tending toward /d/ in terminal and medial positions in contemporary American) and at grammatical (e.g., irregular nouns or verbs) levels. In all such cases, shifts should be in the direction of the stronger (i.e., more frequently used) habits in these divergent hierarchies; /t/ should be under greatest stress to shift toward /d/ in those positions where the total probability of /d/ is higher, and ‘weak’ irregular verbs should be more susceptible to error than ‘strong’ irregular verbs. The latter prediction has been verified for irregular verbs in English, a significant negative correlation being obtained between frequency of errors in writing past tense forms and frequency of usage in Thorndike-Lorge lists.

In the present paradigm it should be noted that the greater the similarity among the divergent responses, the less the interference. This requires some clarification: With increased response similarity there is, to be sure, greater intrusion (substitution) of one response for the other; on the other hand, there is less interference in the sense of blocking, or failure to respond, and in terms of latency of response. This effect is maximal with reciprocally antagonistic reactions, where the learning of one is accompanied by inhibition of the other. This also has interesting implications for language change: In the first place, similar phonemes in terminal positions should tend to merge, e.g., intrude upon each other with such frequency and unpredictability that the difference would lose its distinctiveness. This would help explain the general fact that there is greater diversity in initial position than in terminal position. A similar phenomenon should apply to grammatical affixation—there should be greater diversity among prefixes (and hence differential semantic or lexical significance) than among suffixes (where automatic grammatical significance should be the rule). In the second place, antagonistic reactions in competing divergent positions should tend to interfere with each other in the sense of blocking and increased latency. The combination of these two factors—merger among closely similar reactions and reciprocal blocking among antagonistic—leads to the expectation that the sets of phonemes that appear in given positions following a constant should tend toward an average separation in distinctive features, neither too similar nor too disparate, when frequency is studied.

These psychological considerations offer to shed some light on the typological linguistic problems relating to the contrast between prefixing and suffixing languages. Since inflective and derivational elements are few in number compared to the morpheme membership in root position, the general considerations relating to transitions from divergent hierarchy (in this case root classes) to convergent hierarchies (derivational and inflectional) morpheme classes apply here. In accordance with the principles just discussed, there is greatest facilitation when the divergent hierarchy is followed by the convergent. On this basis, one would expect suffixing languages to be more frequent than prefixing ones, a fact noted by Sapir. One would also expect the prefixing languages to be more fusional (i.e., irregular) and suffixing languages to be more agglutinational (i.e., mechanical) in their morphophonemics. The consideration advanced is that in prefixing languages the difficult transition from a class with few members to a class of wide choices will be ameliorated to the extent that there are special variant forms (alternants) each restricted to a small number of subsequent roots. Indeed, in limiting cases this becomes in a sense a single choice since the enunciation of a prior element which can only be followed by some single subsequent, commits one to this subsequent in advance. The longer this process goes on the more the prior element becomes fused with the subsequent until it ceases to be an independent morpheme.

We might therefore advance the developmental typological thesis that prefixing languages tend towards the isolating type. The evidence both for this and the hypothesis of greater irregularity for prefixing languages needs careful examination before any definitive conclusion can be drawn. Our impression regarding this latter thesis of the more fusional nature of prefixing languages is that it holds in general. Very striking is the total absence to our knowledge of nominal inflectional elements in prefixing position, where the disparity of hierarchies is greatest. In support of the thesis of the development of isolating from prefixing languages, there is striking positive evidence in five cases and no contrary instances of which we are aware. Annamite, Chinese, Thai, Zapotec (Mexico), and Ewe (West Africa) are all classical isolating, monosyllabic languages and all are related to languages of a prefixing type of the Austroasiatic, Sino-Tibetan, Thai-Malayo- Polynesian, Oto-Manguey, and Niger-Congo families respectively.

The empirical fact noted earlier—that the smaller the number of features separating two phones, the greater the probability of change from one to the other—seems to be incorporated in the hierarchy analysis above.

Another principle relating to the locus of formal language change derives from information theory:

V. Sequential redundancy. The more redundant, i.e., predictable, the occurrence of one message element from knowledge of the occurrence of another, the greater the probability of modification of one of them. Information theory in itself does not indicate the precise locus of change, but since messages are unreeled in only one direction, it seems likely that susceptibility to modification should be greater in the terminal members of such redundant sets. This, incidentally, should be a special condition for loss of a phoneme in a language as compared to change. It will probably be necessary to distinguish between inherent redendancy, e.g., where the physiology of the articulatory processes requires it, and incidental redundancy.

(2) Principles Relating to the Process of Change. The process of language change is probably best conceived in terms of the total communication act, involving continuous functional interaction between speakers and hearers. In this process the speaker is the petitioner for changes (cf., Sturtevant’s notion of ‘lapses’) and the hearer is the judge who, via social feedback or differential reinforcement, either allows or refuses to allow each particular modification. Given the existence of many ‘stress points’ in a language, speakers are under more or less continuous, if unconscious, pressures toward modification; to the extent that the process of effective communication in a group is or is not hindered by such modifications, the language will change.

I. Production of changes. Individual speakers of a language will tend to produce changes in those positions and of those types indicated above in proportion to the degree of stress under which they are communicating. By ‘stress’ here is meant any condition of the speaker that reduces his attention to his own self-feedback. Presumably young children in the process of learning language are particularly prone to predictable errors—there seems to be evidence on the rate of change in language suggesting modification in terms of generations of speakers. Similarly, rapid speech, speech under fatigue or any other debilitating condition, speech as it occurs in popular songs, and so forth will be special conditions facilitating change.

II. Social Feedback. To the extent that a speaker modification (a) makes a difference in the code, (b) is not redundant, and (c) occurs in a position of high information value with respect to appropriate behavior, the hearer will differentially reinforce the existing ‘correct’ form. The social relation between the participants in the communication act is also involved here—parents and elders are much more likely to correct children and youths than vice versa, for example. What we are dealing with here are the conditions under which a hearer is likely to notice an erroneous or missing signal and evince this by either a checking verbal response—”What did you say?”—or by unexpected behavior from the speaker’s point of view. If I ask my son for the nail and am handed the pail, to use a crude illustration, I am likety to say, “No, I said nail” with appropriate emphasis and clear articulation of the initial phoneme. Obviously the factors indicated above are interactive: a phoneme shift which makes a difference in the code may nevertheless be passed by the censor if it occurs in a position of low information value in the semantic sense and/or if it is redundant with respect to its phonetic environment (e.g., carried by the allophonic variation in surrounding phonemes). Similarly, a shift may occur in a position of high information value if it does not change the code (or, in doing so, does not produce a different word) and/or if it is in redundant relation to its environment. As young individuals in a language community learn the language as well as the culture, they develop self-correcting tendencies based on self-feedback. However, this self correction is clearly dependent upon the pattern of differential reinforcement received from other members of the community and hence should follow the same principles.

III. Strengthening. Uncorrected modifications, being reinforced as parts of total communicative acts, become stronger habits and compete more and more effectively with ‘correct’ habits. The mere fact of occurrence of changes in predictable loci, e.g., non-random modifications, indicates the existence of underlying readinesses at these points. If we accept the general notion that effective communication is typically rewarding (needed objects are brought to the speaker, his social goals are accomplished, and so forth), then it follows that all stimulus-response sequences contributing to the total communicative act will tend to be reinforced.

IV. Generalization. As the habits producing modifications at ‘stress points’ become stronger, these new response tendencies will generalize or spread to other positions, initially to similar antecedent environments (e.g., similar stimuli) and thence gradually to all environments. This analysis implies that all changes begin their careers as ‘sporadic,’ tending to become ‘conditional’ changes under appropriate conditions, and eventuating as ‘unconditional’ changes. There is considerable doubt among linguists as to the validity of this notion—most apparently assuming ‘sporadic,’ ‘conditional,’ and ‘unconditional’ changes to be different in kind—but the empirical evidence needs to be re-examined from this point of view.⁷¹ The ‘appropriate conditions’ under which sporadic modifications become conditional and conditional changes unconditional are frequency conditions. The present analysis would assume that where changes of a sporadic nature are recorded, the initial occurrence was in a position of maximal ‘stress’ or predictability from both environment and competing habit structure, but the tendency toward generalization was blocked by stronger regular or ‘correct’ habits in other similar environments. Similarly, the restriction of a conditional change to its specific environments suggests that continuing generalization tendencies were blocked by stronger regular habits in other environments in Which the same speech sounds occur. These strong regular or ‘correct’ habits represent ‘frequency mountains’ over which the generalization tendencies cannot pass. Presumably various ‘strong irregulars’ (such as certain nominative plurals like child-children and certain verb forms like go-went) are relics of the language’s past and survive the onslaughts of generalizing changes because of their high frequencies of usage. It also follows from all that has gone before that positions of emphasis—loudness, initial positions of utterances, positions of high information value, and so forth-— will be the most resistant to change since these are positions where modifications produce checking reactions from hearers in the language community.

V. Social change. Language change in a community will be gradual and cumulative, representing a continuous changing proportion of individuals who do or do not hear and produce a particular feature or set of features. The process of change in the community would most probably be represented by an S-curve. The rate of change would probably be slow at first, appearing in the speech of innovators, or more likely young children; become relatively rapid as these young people become the agents of differential reinforcement; and taper off as fewer alnd fewer older and more marginal individuals remain to continue the old fornls. On an empirical level, it should be possible to make a comparative study of forms used as a function of age and other sociological variables. It was suggested that the rate of change may be a function of the size of the language community; it is also undoubedly a function of the status of the communication system. The nature of language change within the individual is a difficult question—some linguists feel that this is an all-or-nothing matter akin to mutation, whereas most psychologists feel that there should be a period, at least, of oscillation between competing forms. Perhaps, in a manner akin to imprinting in birds, individuals never change in the features they hear and produce after early childhood experiences, language changes being purely a matter of sociological shift in the composition of the group. Again, empirical data would have to be collected with this question in mind.

Information theory also generates a very general prediction concerning the direction that language change should take at any given stage. Information theory techniques are readily applicable to phonological changes since we have a unit, the phoneme, which is sharply limited in number for each language and hence susceptible to counting on the basis of texts or lexicon. The same consideration applies to the distinctive features into which phonemes may be analysed. In employing information theory concepts here, two alternatives are available. We may compare two or more stages of the same language attested by written records or we may compare related languages on the basis of assured changes from an ancestral phonological system reconstructed by the well-established techniques of historical linguistics. The comparisons would be in terms of entropy estimates.

Our general hypothesis would be that there are two general factors which influence change iii a phonological system: a tendency toward efficiency and a competing tendency toward redundancy (cf., Zipf’s notions of speaker and hearer economies). A phonemic system may be considered as efficient to the degree that all combinations of distinctive features are utilized in the phonemes, that all phonemes are equi-probable, and that their occurrences are independent of neighboring phonemes. In fact, however, there can never be maximum efficiency—not only would articulatory difficulty of certain combinations of features and of certain sequences always be limiting factors, but channel noise would seriously diminish the comprehensibility of such a system. Perhaps, also, the speed of transmission of semantic units would exceed the channel capacity of the decoder. For these reasons, we may think of a language as maintaining an unstable balance between efficiency and redundancy factors.

VI. Entropy balance. The more the entropy of a given language system deviates from that representing balance between efficiency and redundancy factors the greater will be the tendency to change in the direction of balance. This ‘homeostatic’ principle obviously requires some statement about the balance point. If we give speaker and hearer equal weight in the communication situation, then the balance point would be expected to be 50 per cent efficiency. Synchronically, the calculation of entropy measures of phonological units for an adequate sample of contemporary languages should show something like a normal distribution about this 50 per cent balance point. Diachronically, similarly, a sample of the stages of a given language through time should show a normal distribution about 50 per cent efficiency. Finally, if at a given stage a language is either well above or well below the mean 50 per cent entropy level, we should expect to find either a decrease or an increase in entropy to characterize the subsequent process of change in that language.

A phonemic system may be considered efficient to the degree that the number of distinctive features needed to describe the number of phonemes approaches the minimum of log₂ n, where n symbolizes the number of phonemes in the system. For example, a system of 32 phonemes, using 10 distinctive features, would have an efficiency of exactly 50 per cent, since it could be done with only five distinctive features (log₂ 32 = 5) under conditions of maximum efficiency. We may define the redundancy as 1—E (efficiency). A system of 32 phonemes and 8 distinctive features would be 62.5 per cent efficient and 37.5 per cent redundant. The general lpothesis is that both the average efficiency of different languages synchronically and of the same language diachronically for different stages in its history should approximate the same mean value when this measure is applied.

Unfortunately, very few languages have been analysed in terms of their distinctive features. Nevertheless, the following data display a surprisingly close approximation to predictions: The phonemes of English have been analysed as being 28 in number and requiring 9 binary oppositions.⁷² The efficiency of this system then is 4.80/9 or 53 per cent. Russian phonemes, on the other hand, are 42 in number and employ a total of 11 distinctive features.⁷³ The efficiency of the Russian system is therefore 5.38/11 or 48.9 per cent. The efficiency of modern Spanish, as shown in the table below, is 50.9%. The tendency of three contemporary languages to cluster about 50 per cent efficiency is apparent. The only available presentation of analyses of successive stages of the same language is that of E. Alarcos Llorach for four periods of Spanish.⁷⁴ The inventory of phonemes, distinctive features, and efficiency estimates is given in the following table:

The average of these stages is 51.1 per cent and there is a cyclic trend about this value, both of which are consistent with the hypothesis as applied to diachronic data. Needless to say, no secure conclusions can be drawn from such scanty evidence, but the smallness of the deviations from predictions is striking.

The same prediction, interestingly enough, can be reached on a priori grounds within information theory itself. If we have n distinctive features of binary form, we would expect 2ⁿ possible phonemes. In addition we would expect the selection of a sub-set of phonemes which are maximally discriminable. If the distinctive features were dimensions of a similarity space, we would expect the phoneme regions to be maximally distant (i.e., maximally discriminable). Such a set would be in diagonally opposite regions. If there are n distinctive features, then there would be 2^n–1 maximally distinguishable regions. Then the ratio of existing phonemes to possible phonemes would be

The general notion just described says that languages should tend to change in such ways as to return toward a balance between efficiency and redundancy factors, but it does not specify what types of changes would accomplish these ends. In general, it seems likely that phonemic merger should tend to increase the entropy and hence efficiency of a language system (by reducing the total number of phonemes relative to the same number of distinctive features) while phonemic split should tend to decrease the entropy of a language system. This, of course assumes that the merging phonemes vary in frequency and the result of pooling them will be a more even distribution and hence an increase in absolute entropy. In other words, the preceding analysis assumed approximately equal use of the phonemes in a language. If we had data at our disposal on phoneme sequences, similar predictions could be made with regard to transitional measures, namely, an increase in entropy as a result of merger and a decrease in entropy as a result of splitting.

6.3.1.3. Proposed experiments. A number of research suggestions are embodied in the theoretical analyses above. Additional proposals may be noted here. (1) Prediction of merger and splitting from entropy measures. The information theory analysis just given suggests certain rather obvious tests. For example, measures of relative entropy for Javanese, which has maintained most Proto-Austronesian consonants intact, should be compared with those for Hawaiian or Samoan, in which there has been widespread merger. The process of splitting could be tested by studies of consonant frequencies in the transition from classical Latin to Italian, during which the number of consonant phonemes increased. (2) Re- analysis of historical data. There is an immense amount of evidence available on language change. This should be restudied and sampled in such a way as to test the various hypotheses that have been suggested above. Do we, for example, find that changes tend to occur initially in those positions where anticipatory and perseverative environmental factors combine to modify the intervening phone? Do e find that changes are more probable in terminal and medial positions than in initial? Would a fine enough time series reveal the predicted generalization of changes? (3) Contingency analysis of languages in general. Numerous typologies of world languages have been made, but usually without any clear purpose. The various principles discussed above generate certain predictions as to what language characteristics should (a) appear together and (b) shift from one into the other historically. A matrix in which the columns were defined by characteristics (agglutinative, tonal, etc.) and the rows by a random (or perhaps exhaustive) sample of world languages could be analysed to determine empirically what characteristics tend to appear together and what languages tend to be related to more than chance degrees. This analysis presupposes further development and validation of the specific quantitative indices for each characteristic which have recently been advanced. (4) Experimentally produced lapses. There are a number of ways in which contemporary speakers can be placed under stress and the locus and nature of their lapses recorded and checked against theory. Some of the possibilities are: (a) enforced rapidity of speaking, with tape recording of the results that can then be stretched for analysis; (b) detailed analysis of the spontaneous speech of children of various ages; (c) speech under the conditions of delayed feedback—will the loci of disturbances be predictable from principles such as those above?; (d) analysis of the changes that occur in popular singing; (e) sampling of deliberate ‘humorous’ modifications, e.g., ‘speakers’ to ‘speagers,’ produced from native speakers on request.

6.3.2. Semantic Change

6.3.2.1. Linguistic facts. Change in meanings is as constant a feature of linguistic history as change of forms. Changes in the semantic area of language are in many instances motivated by the introduction of new cultural items or changes in old ones. The basic responses to such situations are usually one of the following: (1) Borrowing from another language. (2) Extension of an old term used to designate something similar either formally or functionally. (3) Coinage of a new term, often by compounding or derivation from previous morphemes. A kind of borrowing in which a new formation is based on the traditional resources of a language but modelled after a formation in a foreign language is sometimes called a caique. The following examples will illustrate these processes. English borrowed street from Latin (via) strata to designate the paved roads which were new to the Anglo-Saxons. It extended the meaning of the already existing verb writan ‘to scratch,’ to include runes on the bark of trees. English formed a new word railroad out of two existing morphemes. An example of caique is the German Fall in the meaning of grammatical case. It is a translation of the Latin casus which has this same meaning in addition to the literal meaning of ‘falling.’ It in turn is a loan translation of the Greek ptosis, literally, ‘a falling.’

Very many changes in meaning can be shown to occur without any precipitating cause in non-linguistic cultural change. These changes like those listed above can be classified as (1) coinages, new forms with new or old meanings whether completely new (noggin) or novel combinations of preexisting elements (big shot). (2) Meaning shifts in preexisting forms. (3) Obsolescence (e.g., loss of the term whither). Meaning shifts can practically all be covered by the term metaphor.’ The various figures of speech of traditional rhetoric, synecdoche, metonymy, etc. name a process by which a term is extended through various associations, e.g., part for whole, whole for part, specialization, generalization, weakening, elevation, degradation, etc. The nonce metaphor of the poet or of the ordinary speaker becomes a meaning shift if it spreads to the rest of the speech- community. If the older meaning becomes obsolescent then a complete shift has taken place, usually with an intermediate period in which both senses exist. More often the various meanings all continue in use, some being viewed as primary others as metaphorical. The prevailing polysemy of words and other linguistic forms is the synchronic result of this diachronic process.

Once a shift has taken place, the result may be a chain reaction. The new meaning of the particular form which has undergone a shift may have been covered by some already existing term. This second term may become obsolescent, or may specialize in some narrower meaning, or in turn shift with further results. For this reason a change in meaning should never be considered in isolation but rather in its effect on a set of forms with related meanings. This is the concept of the semantic field, which has been dealt with especially in Europe.⁷⁵ Specific parallel changes of meaning tend to occur with high frequency in widely separated areas. Such meaning changes can usually occur in either direction. Thus ‘sun’ has become ‘day’ or ‘day’ has shifted to ‘sun’ independently in many languages. Often the same term is used for both, and if historical reconstruction is not possible we cannot tell in which direction the shift occurred. Such instances should be assembled and correlated with word association data (see below).

As in other aspects of historical study we can utilize either historical material proper or that derived from comparison of related languages. A form which is the continuation of an older form in the same language, or shares a common origin with one in a related language as shown by resemblance in form and meaning, is called a cognate. Examples of cognates are (1) Anglo-Saxon stan ‘stone’ and modern English stone in the same meaning. (2) Modern English bone and German Bein ‘leg.’ In this case English has generalized while German keeps the earlier restricted meaning.

6.3.2.2. Theoretical analysis of meaning change. We start with the assumption that ‘semantic change’ in the present context refers to change in the reference of a linguistic sign (or in the semantic state associated with a linguistic sign) and not to change in the significance of referends (objects) themselves as distally perceived. In other words, when the object MOUTH comes to be called mouth rather than cheek, we do not assume that the meaning of this object to human communicators has changed, but rather that the arbitrary linguistic sign by which communicators refer to this object has shifted.

Figure 18

This situation is described in learning theory symbols below and in Figure 18. The distal perceptual sign (₁), deriving from an object (?S₁) elicits its appropriate mediation process (r_m1 —— s_m1) in the encoder, but rather than mediating the encoding of the original linguistic reaction (₁), it now mediates a different linguistic reaction (₂). There is no change in the significance of the perceptual sign (e.g., the meaning of MOUTH), but there is a new encoding unit associated with this object via the mediation process. Similarly, on the decoding side of the communication equation, the original linguistic sign (₁) as a message event no longer elicits a mediation process (r_m1 → s_m1) capable of mediating behavior (R_x1) appropriate to the same object (S₁), but another linguistic sign (₂) does elicit this process. Processes previously associated with the new message event, , if there were such, are indicated by brackets, as are the displaced message events.

Even a casual survey of materials of semantic change as noted above indicates that these shifts are not haphazard. Rather, the referends of the members of a set of cognates tend to be closely related semantically. What general mechanisms of semantic change can be derived from theory which might account for this lawfulness?

(1) Association transfer. The regularities of physics, biology, and culture combine to enforce certain transitional dependencies in sequences of signs. Thus sun to warmth, eye to see, man to work, and vice versa and so forth. As shown in Figure 19, such redundant sequences of signs provide conditions for establishing central associational connections (s_mi → r_m2). Such intra verbal or associative connections are a major determinant of reactions in free association tests. However, as shown by the arrow connecting ₁ with r_m2, this transitional redundancy also provides a condition for a shift in meaning, such that the prior sign comes to signify what was initially an associate. Similarly, on the encoding side, under conditions of high redundancy the representational process associated with the prior sign will tend to become an eliciting condition for encoding the subsequent linguistic unit, as shown by the other arrow connecting s_mi with ₂. This represents a response competition situation in both decoding (competing mediators) and encoding (competing expressions), e.g., divergent hierarchies. To the extent that the mediating reactions r_mi and r_m2 are similar, the subsequent will tend to intrude in place of the antecedent (e.g., if ugly-duckling were a high frequency combination, the heard sign ugly might also tend to elicit the representational process originally associated with the sign duckling, and ugly would be acquiring a different meaning). To the extent that the vocalic expressions ₁ and ₂ are similar, the subsequent will tend to intrude in place of the antecedent (e.g., if bright-light were a high frequency combination, the representational process characteristic of brightness should tend to elicit the encoding unit light—the encoder will have a tendency to substitute the linguistic sign light, as in “It’s a very light day,” in situations appropriate to the original meaning of bright). Both of these processes, mediator substitution and vocalic substitution, have the same end result—a shift in the sign associated with a particular mediation process. Of course, numerous additional factors would be operative in determining the probabilities of such shifts—for example, whether the redundant items are in same or different grammatical classes, relative frequencies of usage, and so forth.

Figure 19

(2) Situational context redundancy. Again by virtue of physical, biological, and cultural regularities, the distal signs associated with certain objects will co- occur with high redundancy. When looking at CHEEK one will nearly always see MOUTH also; when reacting to the COLOR of an object one will also be perceiving its SHAPE. If the male nobility are nearly always seen on horseback, horseman will tend to acquire the meaning of nobility or gentleman (cf., Span, caballero); if the material of which something is made is redundantly and rather exclusively experienced in context with its function, the name for the material may substitute for the function (cf., nickel). As shown below, the greater the frequency of co-occurrence in common situational context of two distal signs, and hence the greater the probability of co-occurrence of their appropriate mediational processes the greater will be the tendency for substitution of one encoding unit for the other (as shown by the dashed arrows). Note that the same shift in encoding units could be accomplished by a change in meaning of the distal signs (e.g., confusion and substitution between r_mi and r_m2), but this need not be the case—the word originally used for JAW can be substituted for that for MOUTH without implying any change in how people perceive these parts of the physiognomy. The likelihood of mediator confusion should depend upon their similarities, and this in turn upon the cultural differentiation between these objects (e.g., to the extent that specific operations, such as chin-beard styling or chin guards for battle, apply to JAW as separate from MOUTH, mediator confusion should be reduced). This distinction between mediator and vocalic skill unit shift verges on the Weltanschauung problem (cf., section 7). It can also be pointed out that situational context redundancy of the sort described here imposes difficulties in the transmission of language from one generation to another, and thus further increases the likelihood of language change—when an adult refers to, and even ‘points at’ his LIPS, the child is quite likely to be reacting in terms of his MOUTH.

Figure 20

(3) Physical stimulus generalization. This is a very straightforward mechanism and simply refers to the tendency for a reaction, in this case mediational, to spread from one stimulus pattern to others, as a function of their physical similarity. Thus the word for SUN should tend to generalize to the object MOON as distally perceived and vice versa; the word for the human EYE should tend to generalize to similar appearing knots in wood, keyholes in doors, and slots in needles; the word for FLOOR should tend to generalize to other level expanses, such as PLAIN and ROAD. The greater the habit strength of a particular mediational and encoding sequence (e.g., the greater its frequency of occurrence), the greater should be its capacity for generalization. Thus, if the name of one carnivorous animal {cat in English) acquires a very high frequency of usage as a result of environmental factors, it may generalize so widely as to substitute for carnivorous animals in general.

(4) Mediated generalization. The mechanism just described is one means whereby a class of stimulus patterns, in that case physically similar, may become, associated with the same label. Mediated generalization is another mechanism but does not require physical similarity. If one object, as distally perceived (e.g., LEG of a person), acquires a certain significance (e.g., ‘stand on’) and another object (e.g., that which holds up a table) acquires through independent conditioning a very similar signification, then the instrumental reactions associated with one sign (in this case, the vocalic encoding unit leg) will tend to transfer via mediation to the other sign. Other illustrations would be the transfer of the term ship to space traveling objects (spaceship) in science fiction, the use of terms like chicken and filly to refer to young girls, and reference to head as the top or most important part of something. Some such process as this presumably is basic to metaphor in general.

6.3.2.3. Research proposals on semantic change. (1) A first step by way of research would be to summarize and categorize (perhaps in terms of the mechanisms described) the various types of semantic changes that have occurred in language families. This should be done across language families as well. Of course, much of this work has already been done and merely needs to be analysed. (2) Given the above collection of information on historical semantic changes, it would be interesting to compare these data with those obtainable from association techniques, using, of course, contemporary subjects. The first mechanism described above, association transfer, involves reactions to linguistic signs, and the usual free association technique should yield data most directly comparable to these types of language change. The other three mechanisms above, situational context redundancy, primary and mediated generalization, involve reaction to the distal or perceptual signs of objects, e.g., labeling operations, and a technique used by Karwoski, Gramlich and Arnott,⁷⁶ among others, where people associated to objects themselves as perceived distally, should yield relevant data. The experimental question is this: to what extent can one demonstrate significant correlation between the associations made to words and objects and evidence on semantic changes that have occurred? Will the types of changes that have been known to occur in languages with considerable frequencies be matched by high frequencies of appearance in association tests? Demonstration of this sort would lend general support to the hypotheses generated here. (3) Study of the verbal paraphasias of aphasics—the encoding units substituted in the search for the ‘correct’ units—would be interesting in its own right as well as instructive in the present context. Do verbal paraphasias, when collected in sufficient numbers, also parallel free association and object association frequencies? (4) The conditions for each of the mechanisms described can be manipulated experimentally under laboratory conditions. For example, one can deliberately vary (a) the sequential redundancy of signs, (b) the contextual (situational) redundancy, (c) the physical similarity of objects, and (d) the development of common significances for dissimilar objects, all in a situation using nonsense materials, and, after training in labeling, measure the errors produced and their natures, the difficulties in training itself, the changes that occur in retention of labels, and so forth.

⁶³ Charles E. Osgood and James J. Jenkins. While members of the seminar were of somewhat different theoretical persuasions, it was agreed that theoretical differences are not critical at this point. This analysis follows Osgood’s mediational theory in the main. For alternative (but not necessarily contradictory) views of language learning cf. B. F. Skinner, Verbal behavior, William James Lectures, Harvard University, 1947, or Roger W. Brown and Don E. Dulaney, A stimulus-response analysis of language and meaning (privately distributed).

⁶⁴ D. 0. Hebb, The organization of behavior (New York, 1949).

⁶⁵ James J. Jenkins and Leonard D. Newmark.

⁶⁶ Susan M. Ervin and Charles E. Osgood.

⁶⁷ Einar Haugen, The Norwegian language in America (1953), and Uriel Weinreich, Languages in contact (Linguistic Circle of New York, 1953).

⁶⁸ Susan M. Ervin.

⁶⁹ Joseph H. Greenberg, Charles E. Osgood, and Sol Saporta.

⁷⁰ See C. E. Osgood, The similarity paradox in human learning: A resolution, Psychological Review 56. 132-143 (1949).

⁷¹ There is also a matter of definition involved here. We are assuming that most ‘sporadic* changes occur in environmental stress points of the sort discussed, while many linguists reserve this term for changes of unlawful, haphazard sorts with respect to environmental factors.

⁷² Jakobson, Fant, and Halle, Preliminaries to speech analysis (Cambridge, 1952).

⁷³ Cherry, Halle, and Jakobson, Toward the logical description of languages in their phonemic aspect, Language 29. 34-46 (1953).

⁷⁴ Fonologia Espanol (Madrid, 1950).

⁷⁵ See particularly Jost Trier, Der deutsche Wortschatz im Sinnbezirk des Verstandes; die Geschichte eines sprachlichen Feldes (Heidelberg, 1931).

⁷⁶ Karwoski, Gramlich, and Amott, Journal of Social Psychology 20. 233-47 (1944).