A core principle of studying language acquisition: It’s a developmental system

This is an exciting time to be studying language development, as is well illustrated in the collection of papers for this special section on core computational principles of language acquisition. These papers highlight innovative approaches that examine the nitty-gritty details of process. In so doing, they reveal a set of fundamental learning processes, but they also emphasize that the subject of our focus is not just a learner and something to be learned. Rather, we seek to understand a complex system in which the information provided is uniquely tuned to the intended recipient who is particularly ready to receive just the bits of information that will carry the system forward and prepare it for the next moment of learning. In particular, these four papers examine critical aspects of the language acquisition system—the input (Hollich and Prince), the nature of the learner (McMurray, Aslin and Toscano), the fit between the learner and input (Chemla, Mintz, Bernal & Christophe), and how learning at one level influences learning at the next (Christiansen, Onnis, and Hockema).

This collection of papers highlight how the complexities of the language input and aspects of the language learner fit together such that the learner can extract patterns at one level and point in time and then submit those extracted patterns to the same types of computations at the next time point to learn even more. Thus, we can see how a simple learner in a structured, dynamic, emergent, and ever evolving context can arrive at something amazing–sophisticated, complex language in 2–3 years! While these papers focus on learning, they also make clear that the history of the system up to the moment of learning make this more than just simple tabula rasa information extraction. These papers show how the complexities of learning—the flow of information through the system, the fit between the learner and the input, and the nonlinearities that the learner imposes on the input conspire to create a process that is much more than learning. This is development (c.f. Spencer, Blumberg, McMurray, Robinson, Samuelson & Tomblin, in press). Recognizing the potential of this kind of developmental explanation for language acquisition comes from an openness to the possibility that emergent, non-linear change can arise from fundamental, basic, unsophisticated principles that combine forming a developmental cascade. This is the promise of this perspective—that greater understanding of one of the most compelling changes in early cognitive development might be understood through the combined power of computational analysis and developmental process.

Hollich and Prince (this volume) focus on the visual input that often accompanies the speech stream. They examine infants’ performance in an audio-visual integration task and ask how much of infant looking behavior in tasks that require this integration can be directly accounted for by a signal-level analysis of the stimuli. Looking behavior is a common measure of early language understanding, but there are complicated issues surrounding how an infant’s looking at (and listening to) a stimulus is related to processing of that stimulus, and how this information processing drives subsequent looking behavior. For example, what causes a release in fixation of one stimulus and a shift to examine another—an endogenous process driven by the amount of information accumulated, an exogenous process driven by salience, or, more likely, both (see Goldberg & Schöner, 2007; Schöner & Thelen, 2006; Perone & Spencer, 2008, March)?

Hollich and Prince begin to examine these issues by comparing the performance of infants to that of models that vary in the particulars of the information they use and the processing they complete on the available information. They find that models that use only signal-level information about visual motion account for many of the details of infant performance. They suggest, therefore, that information contained in the signal presented to infants may serve as a source for audiovisual speech integration by infants. This line of research fits with a number of other recent lines of theory and experimentation that are taking a closer look at the relation between the visual behaviors so commonly used as measures of infant knowledge, and what infants actually process and represent in such tasks (see, for example, McMurray & Aslin, 2004; Perone, Spencer & Schöner, 2007; Schoner & Thelen, 2006). However, Hollich and Prince’s focus on the fit between the audio and visual information presented to children is unique and noteworthy. Thus, while they acknowledge that this work is only a first step in a full understanding of the processes that support infant behavior and does not yet tell us what infants are extracting from the input, further work along these lines will undoubtedly confirm that the combined audiovisual signal is a rich source of information that children can and do use to great advantage.

The paper by McMurray, Aslin, and Toscano (this volume) examines properties of the learner. Specifically they take a critical look at the sufficiency of statistical learning as an explanation of infants’ acquisition of phonetic categories. Using a mixture of Gaussians implementation, they find that, in fact, statistical learning processes alone are insufficient. Competition is needed to insure the correct partitioning of the input space. McMurray et al. then use a full model with competitive processes to explore issues related to the sparseness of the category mapping. This yields four core observations. First, the mechanisms that underlie the observed developmental trajectory are continuious. Second, simply counting the frequency of occurrence of different categorizations is insufficient to produce the categorical distinctions infants make. Third, the process of forming phonetic categories is one of grouping areas of auditory space together, rather than learning particular phonological boundaries. And finally, discrimination between two inputs can be based on a categorization of one input and the lack of a category for another. Thus, behavior that has been previously taken to indicate that infants “have” certain phonetic categories may not actually be based on represented categories in the learner.

By asking critical questions about the nature of the learning processes, McMurray et al. have provided insights to the means by which a key early language-related behavior is realized. Moreover, they have provided an important new perspective on the result of that process. Rather than carving up the phonemic space into fixed categories into which further information will have to be fit, infants are grouping parts of it together, and leaving other parts uncategorized. This has clear implications as the system moves forward and builds on these distinctions because it is left in a more flexible state ready to incorporate future input.

The exquisite fit between what the learner needs to acquire and what she has available to do the learning is well illustrated in the work of Chemla, Mintz, Bernal and Christophe (this volume). These authors suggest how infants could extract higher-order categorical information via simple computations on lower level structure. They propose that the lexical category of target words can be determined by tracking the co-occurrence of two context words that surrounded the target. Their first experiment extends this “frequent frame” approach, which was previously developed on a corpus of English infant-directed speech, to French. Chemla et al. show that while French may appear to be problematic for the frequent fames approach, it is, in fact, completely amenable. A second experiment shows that not just any pair of co-occurring words near a target can be used for categorization. Rather, frames need to surround target words. A third experiment demonstrates that frames are item-specific; recursion of frames does not produce even better categorization.

Together, Chemla et al.’s experiments show that if an infant was tracking information about individual words, the structure of the input could produce an anchor by which that child could extract higher-level information about the words, without having to shift to explicit computations at that higher level. Moreover, this work shows that not only do we not need to build in knowledge of lexical categories prior to learning, it is actually better for infants to start off without categories. Infants can track information about something we know they have access to early—individual words (see Christiansen et al., this volume, for example)—and get to something more—categorical structure. In this way, then, the abilities of the learner are well suited to the information provided.

Christiansen, Onnis and Hockema (this volume) provide another way in which infants could extract higher-order categorical information via simple computations on lower level structure, and show how this process can go one step further. These authors provide a computational analysis of English child-directed speech that progresses over two stages. First, transition probabilities between phonemes were used to find words in a stream of unsegmented speech. Second, distributional information about initial and final phonemes is used to predict the syntactic category of words isolated in the first stage. Christiansen et al. suggest that a core computational principle of language acquisition is that the same source of information can be used to learn about multiple aspects of language structure—words and lexical categories. Moreover, this work elegantly illustrates how a very simple computational process—tracking patterns in the stream of continuous sounds—can move the learner to a higher-level understanding of structure, without having to build the resultant categories into the system. Thus, by performing the same simple computations on the successively more structured product of its own computations, even a relatively unsophisticated learner can capture what has previously been seen as unobtainable abstract structure.

As is clear in these papers, the field of language development has developed a greater appreciation for infants’ and toddlers’ amazing language learning abilities. This combined with the increasing use of innovative and quite sophisticated modeling techniques has driven a shift in focus away from descriptions of language behaviors and abilities at different developmental stages, and towards the details of developmental process. Increasingly, the central question is what are the mechanisms of change. And in answering this question there is a corresponding move to focus on the nature, structure and richness of the input to learners at multiple timescales and very fine-grained levels of detail. This is equally complemented by a greater understanding of the learner as situated in a rich and supportive context that evolves over time. Thus, when the field previously asked “what information do language learners have to work with?” and “what can they learn with that information?” the answers were “too little” and “not much.” More recent work such as that presented in this collection, however, suggests the answers to these questions are actually, “a lot” and “quite a bit”. An important question for future work, then, will be what are the processes by which this evolution from learning to development occurs. That is, how individual instances of in-the-moment learning coalesce and build on a longer timescale into developmental change (see also McMurry, Toscano, Horst & Samuelson, in press; Samuelson & Horst, 2008). Understanding the core computational principles of language development is an important step in this direction.