The problem of language acquisition

As mentioned above, there are no pills or injections that the traveller could be given to provide her/him with sufficient knowledge of Chinese when travelling to China, of Finnish when travelling to Finland, and so forth. Despite the specific human endowment to acquire language, we experience growing difficulties to acquire languages as we grow in age. This simple observation has led cognitive scientists to extricate themselves from classical learning theories. Indeed, why is it that at an age when the ability to learn mathematics or spelling is still absent, children are so proficient in learning a second or even a third language? And why is it that a few years later, younger adults will learn as many topics as needed while in college but will display severe problems when they have to acquire a new language? Even the sarcastic comments of Voltaire do not clarify matters. In his Micromegas,⁹ he imagines a Cartesian thinker saying: ‘L’âme est un esprit pur, qui a reçu dans le ventre de sa mère toutes les idées méthaphysiques, et qui, en sortant de là est obligé d’aller à l’école, et d’apprendre tout de nouveau ce qu’elle a si bien su et qu’elle ne saura plus (The soul is a pure spirit which receives in its mother’s womb all metaphysical ideas and which on issuing thence, is obliged to go to school as it were and learn afresh all it knew so well, and will never know again). In fact, cognitive scientists moved away from the traditional association learning theories, and started favouring species-specific learning mechanisms, even though these do not remove or solve all the problems either, as Voltaire’s astute and ironic comment suggests.

In this chapter our purpose is to present some language learning mechanisms, which are the object of very active experimental investigations. In fact, Chomsky’s theoretical arguments persuaded many cognitive scientists to explore the notion that only the human mind comes equipped with a specialized Language Acquisition Device (LAD). Chomsky (1965,Reference Chomsky¹⁰ 1980;Reference Chomsky¹¹ see also Rizzi 1986Reference Rizzi¹²) proposed a theoretical framework called the Principle and Parameters (P&P) approach, which postulates universal ‘principles’ that apply to all natural languages, and a set of binary ‘parameters’ that have to be set according to the grammar of the surrounding language. Principles are mandatory for all natural languages, whereas parameters constrain the possible format of binary properties that distinguish groups of natural languages. The P&P formulation helped cognitive science propose realistic scenarios to explain how humans overcome the difficulties of first language learning.

Chomsky recently wroteReference Chomsky¹³ that

Further along Chomsky states that

The LAD would allow infants to set the different grammatical parameters to the value of their language of exposure. The proposal had a great influence on the field of language acquisition. Over the last two decades, however, it has become apparent that even if the LAD is a very attractive formal description of language acquisition, it does not specify in detail what it is that counts as a parameter, or even how many parameters need to be set in order to acquire a language. Nor did this formal proposal clarify how parameters are actually set given exposure to the linguistic data the learner receives.

Generative theorists have pointed out that the linguistic environment is far too impoverished for infants to learn the grammar of the native language on the basis of classical learning mechanisms. In fact, classical learning theories do not postulate a LAD or a specialized mechanism that would explain why it is that only humans can acquire grammatical systems. Moreover, to explain how infants acquire the grammar of their language of exposure, given that today over 6000 languages are spoken in different parts of the world, it is necessary to elucidate why it is so easy for infants to learn with equal facility any one of the languages that happens to be spoken in its surroundings. Biological observation suggests that the human species is endowed with special mechanisms to learn language. We thus need to understand how these putative mechanisms are deployed, how they interact with one another, and the order in which they are engaged. In other words, the experimental study of early acquisition has to be recognized as an essential area to judge theories for their formal pertinence and for biological validity.

Experimental studies show that considerable learning occurs during the first months of life (see, among others, Refs Reference Kuhl and Williams14–Reference Morgan18). Moreover, a number of investigators began to defend the notion of phonological and prosodic bootstrapping, that is, the notion that infants might explore the cues in speech that correlate with abstract properties of grammar (see, among others, Refs Reference Wanner and Gleitman19–Reference Cutler and Mehler23). The multiple cues that transpired from these studies make it possible to understand how and why it is that the notion expressed in the P&P approach might receive empirical support. In brief, research with very young infants supports the view that the signals received during the first year of life are potentially very rich in information and that the human brain, endowed as it is with mechanisms to extract the information carried by speech, will use them to discover abstract grammatical properties.

Below, we lay out in some detail some of the specific mechanisms that are used during the first years of life. First, we focus upon the faculty of the infant’s brain to compute distributional properties of speech signals. Next, we explore the ability of the human mind to project generalizations on the basis of very sparse data. Finally, we explore some of the constraints that underlie both of the above mechanisms and how they handshake when they operate simultaneously.

Distributional computations

It has long been recognized that speech utterances tend to display statistical dependencies between phonological or lexical categories; some of those dependencies are verified in all spoken languages while others are language-specific.Reference Zipf^24,Reference Miller²⁵ Hayes and ClarkReference Hayes and Clark²⁶ demonstrated that, in languages with multi-syllabic words, the transition probability (TPs) between word internal syllables tends to be higher than the TPs between the last syllable of any word and the first syllable of the next one. At first, this demonstration failed to attract much attention. Cognitive scientists seemed to realize its potential importance only after Saffran et al.’s workReference Saffran and Aslin²⁷ was published. Indeed, the results of Saffran et al. fulfilled the long-standing hope of discovering an unlearned neural computation that could provide insight into the ways in which very young infants manage to break into the constituents of their mother tongue. Consider how infants learn words from the speech signals in their environment. Most of the utterances are continuous. It is sufficient to look at the waveform of the utterances to be convinced that there are hardly any pauses, and that the few that there might be do not necessarily signal words. As adults we almost never find ourselves in a situation in which we have difficulties parsing utterances we hear – of course, utterances that belong to our native language – and thus we tend to think that infants also have an easy time parsing the utterances they hear. This is, however, not the case. Before babies are able to recognize a few words, they need to listen to utterances from the language spoken in their milieu, for almost a year. How then, do infants segregate the words that are contained in the utterances to which they are exposed? One possible suggestion is to claim that some words are presented in isolation. This, however, is a rare phenomenon.

Saffran et al.Reference Saffran and Aslin²⁷ discovered that eight-month-old infants can segment artificial continuous streams of syllables assembled in such a way that there are tri-syllabic ‘words’ that have high TPs between adjacent syllables while between words the TPs are lower. In the experiment, four ‘words’ were used; each word contained syllables of equal duration and intensity. Infants were familiarized with a stream for two minutes, and then were tested to establish whether they had extracted the ‘words’ on the basis of TPs. If they had, they should react differently to ‘words’ and to ‘part-words’, which contain the two last syllables of a word and the first syllable of another word, or the last syllable of a word and the first two syllables of another word. The results show that infants do react differently to ‘part-words’ and to ‘words’ in a head-turning paradigm.

Saffran et al.Reference Saffran and Aslin²⁷ convincingly demonstrated, then, that TPs, in the absence of any other cues, make it possible to parse such speech streams. We shall argue below that distributional cues are a very powerful kind of computation that can be exploited on line for segmentation as well for the extraction of other properties of the speech signal. In a recent publication, Gervain et al.Reference Gervain and Nespor²⁸ studied how infants could gain an understanding of the word order properties of their mother tongue before they had access to lexical items. The authors show that eight-months-old Japanese and Italian infants have opposite ‘word-order’ preferences after having been exposed to an artificial grammar continuous string composed of high frequency words and low frequency words in alternation, with each word being monosyllabic. After listening to the speech stream, Japanese and Italian infants were confronted with two-syllable items: one item had a high frequency syllable followed by a low frequency syllable; the other item had the two syllables in opposite order. Japanese and Italian infants displayed a preference for different test items. This suggests that infants possess some representation of word order prelexically. The authors propose a frequency based bootstrapping mechanism to account for the above results. They argue that infants might build representations by tracking the order of function and content words, identified through their different frequency distributions.

Despite the power of distributional cues that are present in artificial speech, it is necessary to pursue such studies with methodologies that attempt to use strings that resemble more closely the utterances that the infants are processing in more realistic interactions. For instance, Shukla et al.Reference Shukla and Nespor²⁹ used an innovative design to have as much control of the stimuli as in an artificial grammar while adding prosodic cues to it. Shukla et al.Reference Shukla and Nespor²⁹ studied in greater detail how prosodic structures, and in particular intonational phrases (IP), interact with TP computations using an artificial speech stream. The IPs were modelled from natural Italian IPs. They found that when a ‘statistical word’ appears inside an IP, participants recall it without problems. This is not observed when one syllable of the same ‘statistical word’ appears at the end of an IP while the other two syllables appear in the next IP. Moreover, it is unimportant whether the Italian IPs are substituted by Japanese IPs, showing that the primacy of prosody over TPs is not necessarily the result of experience with one’s native language.

Even while recognizing that the brain of very young infants has the capacity to track distributional properties present in signals, we still need to question whether distributional computations are sufficient to do away with much of the Chomskyan generative grammar notions, as was claimed by Bates and Elman.Reference Bates and Elman³⁰ Indeed, each infant has to learn her/his native language; the human brain must have the power to acquire the knowledge to produce and to understand any grammatical utterance of the language of exposure, a learning ability that no other animal brain displays. Even though animals have been shown to compute TPs, much as human infants do, they do not learn language.Reference Hauser and Newport³¹ Hence, although we do not question the importance of statistics, we ask which other mechanisms are necessary to attain grammar and how do they interact with one another.

Before we move to the next section we want to remind the reader that in natural languages non-adjacent dependencies hold between words as well as between morphological components. In the next section, we present experimental work showing that participants use non-adjacent TPs to parse continuous speech streams; we also present evidence that under some stimulating conditions participants discover and generalize to novel instances on the basis of regularities that were present during the familiarization phase. We will argue that the ability to project conjectures is another mechanism that is used in all kinds of situations and, in particular for extracting some grammatical properties, is a most singular accomplishment of the human mind.

Rules and generalizations

Peña et al.Reference Peña and Bonatti³² set out to explore experimentally how the two mechanisms mentioned above (namely statistical computations and the projection of conjectures) interact with one another. In their first experiment the authors presented a continuous Saffran et al.-like stream to participants. The stream had high TPs between the first A syllable and the last C syllable of A x C like words, while the x varied. Participants were able to use the non-adjacent TPs to segment the stream into the A x C words, even though the xs of the test items had never occurred between an A and a C during familiarization. In their next experiment, the authors investigated whether participants could also extract the underlying regularity governing the ‘words’, namely that the first syllable A_i predicts the last one C_i regardless of the syllable(s) inserted between them. The first experiment evaluated the performance of the participants with a two-alternative forced choice test between a ‘word’ and a ‘part-word’ (a ‘part word’ can take two forms, namely CAX or XCA). The results demonstrated that participants chose words as being more familiar than part-words. In the second experiment, when participants were confronted with a ‘part-word’ versus a ‘rule-word’ (a ‘rule-word’ has the form AX*C where X* is a syllable that never appears in that position during familiarization) they failed to show any preference for one of the test items. Therefore, when continuous monotonous streams are used during familiarization, participants fail to generalize a regularity such as ‘if A in first position then C in last position regardless of which syllable appears in between them’ although they can extract ‘words’ from the same streams relying on TPs between non-adjacent syllables.

Obviously, under more naturalistic situations, language learning infants never confront utterances composed of syllables of equal duration, equal intensity and with no pauses or modulations between the ‘words’. Universally, speech utterances have complex prosodic structures. Although there are some prosodic variations between languages, what is never observed in well-formed utterances is the flat, isochronous syllables without pauses used in artificial grammar experiments. Consequently, we ought to try to establish how the brain operates when confronted with equally well-controlled stimuli when prosodic-like cues are inserted. For instance, how would participants behave if the continuity of the streams that Peña et al.Reference Peña and Bonatti³² used in the above experiments is interrupted? Could a minor change in the stimulating conditions change the mechanisms the brain implements to extract information during the familiarization?

We hypothesized that when the brain is confronted with streams such as those used by Saffran and her colleagues, it will exploit the only mechanism that is left to try, namely computing distributional regularities to introduce, possibly, some structure into the otherwise homogeneous speech signal. Instead, when minimal pauses are introduced, participants may be able to learn that words start and end with predictable syllables. As we will see, generalizations are usually projected on the basis of sparse data and not through purely statistical computations. Peña et al.Reference Peña and Bonatti³² introduced silent pauses of 25 ms between the statistical ‘words’, which were tested in the two experiments described above. The very short pauses we inserted were not salient, and generally participants reported not hearing pauses when they were questioned at the end of the experiment. However, their processing mechanisms obviously were sensitive to the pauses, since the manipulation proved to be sufficient for participants to judge that ‘rule-words’ were far more familiar than ‘part-words’. Even when the familiarization was reduced from 10 to 2 min, participants continued to judge ‘rule-words’ as more familiar, suggesting that mechanisms that verify regularities operate on the basis of very sparse data.

In an earlier experiment Marcus et al.Reference Marcus and Vijayan³³ made a point that was, in part, close to one of Peña et al.’s points.Reference Peña and Bonatti³² In fact, Marcus et al. were also persuaded that very young infants are capable of drawing generalizations on the basis of data short habituation. In fact, these authors proposed that to explain the acquisition of linguistic structures, next to simple distributional mechanisms, other mechanisms are indispensable. In particular, they argued that 7-month-olds use algebraic-like rules to extract and encode grammatical regularities. In order to support their claims, Marcus et al. carried out several experiments that were extremely influential because they suggested that 7-month-olds familiarized with structured items were able to generalize the structure to entirely novel items. For instance, in one experiment they familiarized the infants with AAB structured items (e.g. fifigo, lalaru, etc) and during the test phase the participants preferred A*A*B* items (where the * indicates syllables that had never been used during the familiarization) over A*B*B* items (gofifi, rulala, etc). The authors explain the infants’ preference as arising because the participants notice that the ABB underlying structure differs from the AAB used during familiarization. Moreover, Marcus et al. proposed that statistical machines cannot account for the extraction and generalizations of structural regularities between some syllables, when the test concerns novel syllables. Even if different non-statistical mechanisms may be invoked, as we will see below,Reference Endress and Scholl³⁴ it is clear that infants at some point in life begin making powerful generalizations. In fact, Marcus, like ChomskyReference Chomsky³⁵ and Fodor³⁶ amongst others, holds the view that one of the most characteristic abilities of our species is to conjecture rules to learn regularities present in the input data we receive. Some of the authors cited suggest that rule-governed behaviour is at the basis of language acquisition.

Recently, Gervain et al.³⁷ explored the behaviour of newborn infants when confronted with items that share an ABB structure. They used a near-infrared spectroscopy device (NIRS) that allows observation of stimulus-related activation in 24 areas of the cortex, 12 over the left and 12 over the right perisylvian areas. The authors compared the blocks when infants listen to items whose first syllable differs from the second one, which reduplicates in third position (ABB), with interleaved blocks during which the newborns listened to items whose three syllables differed from one another (ABC). The activations in response to the ABB items differ significantly from those the cortex of the newborns displays when confronted with the ABC items. The observed differences make it clear that the Marcus ABB-kind of regularity is processed very differently from the way an ABC pattern is processed. While the ABB items have a structural property that is easy to extract, the ABC items either do not give rise to a common structure (i.e. the regularity common to barisu, fekumo, etc), or it gives rise to many different structural properties, e.g. all items are tri-syllabic or all items have only different syllables, or all items have only CV syllables, etc. Newborns are not very apt at counting or having meta-linguistic abilities, nor do they formulate easily conjectures about what a series of items do not have in common. In contrast, the brain of the neonate, much like the nervous system of other animal species, has the ability to detect adjacent repetitions and to generalize from a series of items to novel ones.

Gervain et al.³⁷ also noticed that the newborns’ cortical responses were different at the beginning compared with the end of the experiment. In fact, a time-course analysis indicated that the responses to the ABB items increase over time in some regions of interest, whereas response to the ABC items remains flat throughout the duration of the experiment. Taken together, these results suggest that newborn infants can extract the underlying regularity of the ABB items even though there are many different syllables that were used to assemble the individual items. Of course, there are, as we mentioned above, two possible mechanisms. The first postulates the computation of algebraic-like rules, such as ‘all items have one syllable followed by a different one that duplicates’. The second is a mechanism to detect repetitions that, at least for the auditory modality, generalizes a structural property to items that repeat adjacently regardless of the token syllables that are used to implement the repetitions.

In a second experiment very similar to that described above, Gervain et al.³⁷ showed that the results were very different when the test items included only non-adjacent repetitions, i.e. items that conform to the ABA pattern, versus ABC items. The results indicate that there is no effect due to the extraction of an algebraic-like rule in this case. Nor is there either a hemispheric asymmetry effect or a change of activation pattern during the experiment. Indeed, when a time course analysis was carried out, the authors failed to find any effects. These results suggest that, at the initial state, a repetition detector fails to respond when repetitions are non-adjacent. This might signify that the repetition detection mechanism is not functional when there is a time-gap between the two A items that is greater than a few hundredth of a second. Further studies are being carried out to understand exactly the mechanism underlying the response to repetitions in very young infants.

Perceptual primitives, memory effects and other constraints

Over the last few years, an interesting discussion has arisen about the likelihood that the abilities of infants to acquire abstract knowledge are uniquely or mostly driven by algebraic-like rules. Obviously, focusing on the ability of using a mechanism of generalization that operates using propositional attitudes and algebraic-like rules seems to us to be an important move. It is unimaginable to hold the opposite view. How could science, the arts, politics and civilization emerge otherwise? Likewise, it seems most unlikely to view all acquisition of knowledge to be the result of just statistical learning of which associative learning is only one aspect. There is sufficient evidence showing that there is ample opportunity to extract and use such information during the course of language acquisition. However, language and knowledge acquisition are, in all likelihood, relying on several mechanisms that function either in parallel or during some time-locked phases of acquisition. Moreover, as we argued before, it appears that generalizations are drawn on the basis of very meagre information while the extraction of statistical information requires a greater amount of information. Likewise, there are other properties of the cognitive apparatus that influence the nature and the outcome of language acquisition. We call processing constraints those constraints that arise from the very nature of some perceptual and memory primitive processing operations.

It is clear that basic cognitive processes interact with language acquisition. It would thus be very surprising if such processes had failed to influence basic language acquisition processes. In fact, Endress et al.Reference Endress and Scholl³⁴ demonstrate that only repetition-based structures with repetitions at the edges of sequences (e.g. ABCDEFF but not ABCDDEF) can be reliably generalized, although token repetitions can easily be discriminated at both sequence edges and sequence middles. These results suggest that there are interesting positional constraints that license spontaneous generalizations at certain positions of a speech stream but not at others. Why should that be? One answer to this question may reside in the fact that only the edges of a speech stream are salient, possibly an auditory ‘gestalt-like’ organization that has to be explained on a par with any other gestalt phenomenon. Otherwise, it is likely that only the edge positions of the stream count as true variables to which values can be assigned. The values of the other positions would refer to the initial and the final syllables.Reference Burgess and Hitch³⁸

In several experiments, further support was gathered for such a view. For instance Endress and Mehler (in preparation) demonstrate that it is possible to use penta-syllabic items (A_iXYZE_i) to extract generalizations, namely to recognize that (A_kX′Y′Z′E_j) is accepted as a familiar item even it was never presented. This occurs only when each separate item is presented individually but not when a continuous stream is used. Moreover, when the second and fourth positions were used (as in XA_iYE_iZ ) participants were unable to generalize when items were presented individually. Nevertheless, they were able to use positions 2 and 4 to compute transition probabilities to segment a continuous stream. These results might not only clarify what participants do when confronted with artificial grammar learning experiments. They may also help explain why it is that most languages of the world use pre- and suf- fixes rather than infixes. Indeed, since edges of content words are more salient, it is reasonable to expect those positions to be privileged for inserting morphological markers. In fact, Endress and BonattiReference Endress and Bonatti³⁹ have described certain phenomena that only occur at the edges of items like those used in Peña et al.’s experiments.Reference Peña and Bonatti³² Indeed, when participants are confronted with AxC items in non-continuous streams they tend to learn that there is a class of A items and another class of C items that can pair in ways that did not arise during familiarization, e.g. any A_i item can pair with any C_k item.

The above experiments ought to satisfy even the most critical readers of the existence of constraints that make edges very salient and highlight their role as variables, whereas the distance from the nearest edge is used to tag internal positions. It also shows that while TPs can be computed in any position, it is only in edges that generalizations can be projected. We suggest that a statistical machine cannot project generalizations and thus cannot explain these results.

Humans display a powerful perceptual mechanism derived from biological movement, namely, the iambic-trochaic law described in Nespor et al.⁴⁰ The signal determines grouping at the phrasal level and may thus be a cue to the value of the head-complement parameter: if phonological phrase stress is realized mainly by pitch and intensity (as in Turkish), then words are grouped trochaically, if mainly through duration and intensity (as in French) then they are grouped iambically.

Since in a pair of words of which one is the complement and one is the head, stress always falls on the complement, the particular physical manifestation of stress might indicate to the prelexical infant the relative order of head and complements. The same physical correlates of stress are found also intralinguistically: in German, the phonological phrase stress of a complement preceding its head is realized mainly through pitch and intensity, of a complement following its head mainly through duration.

Originally, the iambictrochaic law, proposed both for musicReference Bolton^41, Reference Woodrow and Stevens⁴² and for low levels of linguistic structure,Reference Hayes⁴³ was based on intensity and duration only. Bion et al.⁴⁴ conducted a series of experiments to verify the role of pitch in grouping: do adults as well as infants indeed create prominent initial chunks if prominence is realized through pitch, and in prominent final chunks, if prominence is realized through duration? Similar experiments were also conducted on visual images and on gestures. All confirmed the hypothesized grouping. This suggests that the iambic-trochaic law is a further constraint based on perception across modalities and derived from biological movement: high frequency and short duration characterize the onset of groups and low frequency and longer duration their end.

In all the above examples of the primitives and the constraints we individualized, it appears that basic cognitive processes, e.g. memory, perception, gestalt-like emergents and biological movement are at the origin of their existence. These specialized tools modify dramatically the standard computational process displayed in higher mental processes and may help understand how the human faculties take the form they display. For instance, natural languages as well as logic may be the consequence of the interactions of statistics, rule-guided behaviour, and the primitive tools described above.