Language acquisition in the absence of explicit negative evidence: how important is starting small?

Artificial Neural Network Language Models Predict Human Brain Responses to Language Even After a Developmentally Realistic Amount of Training

Artificial neural networks have emerged as computationally plausible models of human language processing. A major criticism of these models is that the amount of training data they receive far exceeds that of humans during language learning. Here, we use two complementary approaches to ask how the models' ability to capture human fMRI responses to sentences is affected by the amount of training data. First, we evaluate GPT-2 models trained on 1 million, 10 million, 100 million, or 1 billion words against an fMRI benchmark. We consider the 100-million-word model to be developmentally plausible in terms of the amount of training data given that this amount is similar to what children are estimated to be exposed to during the first 10 years of life. Second, we test the performance of a GPT-2 model trained on a 9-billion-token dataset to reach state-of-the-art next-word prediction performance on the human benchmark at different stages during training. Across both approaches, we find that (i) the models trained on a developmentally plausible amount of data already achieve near-maximal performance in capturing fMRI responses to sentences. Further, (ii) lower perplexity-a measure of next-word prediction performance-is associated with stronger alignment with human data, suggesting that models that have received enough training to achieve sufficiently high next-word prediction performance also acquire representations of sentences that are predictive of human fMRI responses. In tandem, these findings establish that although some training is necessary for the models' predictive ability, a developmentally realistic amount of training (∼100 million words) may suffice.

Predictability and Variation in Language Are Differentially Affected by Learning and Production

General principles of human cognition can help to explain why languages are more likely to have certain characteristics than others: structures that are difficult to process or produce will tend to be lost over time. One aspect of cognition that is implicated in language use is working memory-the component of short-term memory used for temporary storage and manipulation of information. In this study, we consider the relationship between working memory and regularization of linguistic variation. Regularization is a well-documented process whereby languages become less variable (on some dimension) over time. This process has been argued to be driven by the behavior of individual language users, but the specific mechanism is not agreed upon. Here, we use an artificial language learning experiment to investigate whether limitations in working memory during either language learning or language production drive regularization behavior. We find that taxing working memory during production results in the loss of all types of variation, but the process by which random variation becomes more predictable is better explained by learning biases. A computational model offers a potential explanation for the production effect using a simple self-priming mechanism.

Linking Language to Action: Enhancing Preschoolers' Communicative Abilities Within Language Stimulation

PURPOSE

Children's vocabulary and syntactic skills vary upon school entry in depth and breadth, persistently influencing academic performance, including reading. Enhancing early communicative abilities through multisensory, playful, and conversational experiences is essential and will benefit children's school readiness. This study investigated whether a language-to-action link created during language stimulation, which combines multisensory input, play, and conversation using clay, improves preschoolers' communicative abilities in terms of vocabulary, syntactic, and pragmatic language abilities more than traditional toy-based language stimulation.

METHOD

Language skills were examined in a pre- to posttest design in which 43 typically developing participants, ages 3-5 years, were randomly assigned to clay-based (n = 24) or traditional play-based (n = 19) language stimulation for 8 weeks.

RESULTS

Receptive and expressive vocabulary knowledge for items introduced in the language stimulation program, mean length of utterance (MLU), and conversational initiations improved for participants in the clay condition, whereas significant language skill growth was not observed for participants in the traditional play-based stimulation condition with toys.

CONCLUSIONS

A language-to-action link is created when children engage with open-ended materials, such as clay, as they craft target objects hands on and step by step, affording additional opportunities for language input and output. Results preliminarily suggest that using open-ended materials may enhance children's communicative abilities in receptive and expressive vocabulary, syntax/MLU, and pragmatics (i.e., conversational initiations) more than prefabricated toy objects during language stimulation.

SUPPLEMENTAL MATERIAL

https://doi.org/10.23641/asha.24093780.

A Survey on Curriculum Learning

Curriculum learning (CL) is a training strategy that trains a machine learning model from easier data to harder data, which imitates the meaningful learning order in human curricula. As an easy-to-use plug-in, the CL strategy has demonstrated its power in improving the generalization capacity and convergence rate of various models in a wide range of scenarios such as computer vision and natural language processing etc. In this survey article, we comprehensively review CL from various aspects including motivations, definitions, theories, and applications. We discuss works on curriculum learning within a general CL framework, elaborating on how to design a manually predefined curriculum or an automatic curriculum. In particular, we summarize existing CL designs based on the general framework of Difficulty Measurer + Training Scheduler and further categorize the methodologies for automatic CL into four groups, i.e., Self-paced Learning, Transfer Teacher, RL Teacher, and Other Automatic CL. We also analyze principles to select different CL designs that may benefit practical applications. Finally, we present our insights on the relationships connecting CL and other machine learning concepts including transfer learning, meta-learning, continual learning and active learning, etc., then point out challenges in CL as well as potential future research directions deserving further investigations.

The neural architecture of language: Integrative modeling converges on predictive processing

The neuroscience of perception has recently been revolutionized with an integrative modeling approach in which computation, brain function, and behavior are linked across many datasets and many computational models. By revealing trends across models, this approach yields novel insights into cognitive and neural mechanisms in the target domain. We here present a systematic study taking this approach to higher-level cognition: human language processing, our species' signature cognitive skill. We find that the most powerful "transformer" models predict nearly 100% of explainable variance in neural responses to sentences and generalize across different datasets and imaging modalities (functional MRI and electrocorticography). Models' neural fits ("brain score") and fits to behavioral responses are both strongly correlated with model accuracy on the next-word prediction task (but not other language tasks). Model architecture appears to substantially contribute to neural fit. These results provide computationally explicit evidence that predictive processing fundamentally shapes the language comprehension mechanisms in the human brain.

Input Complexity Affects Long-Term Retention of Statistically Learned Regularities in an Artificial Language Learning Task

Statistical learning (SL) involving sensitivity to distributional regularities in the environment has been suggested to be an important factor in many aspects of cognition, including language. However, the degree to which statistically-learned information is retained over time is not well understood. To establish whether or not learners are able to preserve such regularities over time, we examined performance on an artificial second language learning task both immediately after training and also at a follow-up session 2 weeks later. Participants were exposed to an artificial language (Brocanto2), half of them receiving simplified training items in which only 20% of sequences contained complex structures, whereas the other half were exposed to a training set in which 80% of the items were composed of complex sequences. Overall, participants showed signs of learning at the first session and retention at the second, but the degree of learning was affected by the nature of the training they received. Participants exposed to the simplified input outperformed those in the more complex training condition. A GLMM was used to model the relationship between stimulus properties and participants' endorsement strategies across both sessions. The results indicate that participants in the complex training condition relied more on an item's chunk strength than those in the simple training condition. Taken together, this set of findings shows that statistically learned regularities are retained over the course of 2 weeks. The results also demonstrate that training on input featuring simple items leads to improved learning and retention of grammatical regularities.

Object-Label-Order Effect When Learning From an Inconsistent Source

Learning in natural environments is often characterized by a degree of inconsistency from an input. These inconsistencies occur, for example, when learning from more than one source, or when the presence of environmental noise distorts incoming information; as a result, the task faced by the learner becomes ambiguous. In this study, we investigate how learners handle such situations. We focus on the setting where a learner receives and processes a sequence of utterances to master associations between objects and their labels, where the source is inconsistent by design: It uses both "correct" and "incorrect" object-label pairings. We hypothesize that depending on the order of presentation, the result of the learning may be different. To this end, we consider two types of symbolic learning procedures: the Object-Label (OL) and the Label-Object (LO) process. In the OL process, the learner is first exposed to the object, and then the label. In the LO process, this order is reversed. We perform experiments with human subjects, and also construct a computational model that is based on a nonlinear stochastic reinforcement learning algorithm. It is observed experimentally that OL learners are generally better at processing inconsistent input compared to LO learners. We show that the patterns observed in the learning experiments can be reproduced in the simulations if the model includes (a) an ability to regularize the input (and also to do the opposite, i.e., undermatch) and (b) an ability to take account of implicit negative evidence (i.e., interactions among different objects/labels). The model suggests that while both types of learners utilize implicit negative evidence in a similar way, there is a difference in regularization patterns: OL learners regularize the input, whereas LO learners undermatch. As a result, OL learners are able to form a more consistent system of image-utterance associations, despite the ambiguous learning task.

A failure to replicate rapid syntactic adaptation in comprehension

Language comprehension requires successfully navigating linguistic variability. One hypothesis for how listeners manage variability is that they rapidly update their expectations of likely linguistic events in new contexts. This process, called adaptation, allows listeners to better predict the upcoming linguistic input. In previous work, Fine, Jaeger, Farmer, and Qian (PLoS ONE, 8, e77661, 2013) found evidence for syntactic adaptation. Subjects repeatedly encountered sentences in which a verb was temporarily ambiguous between main verb (MV) and reduced relative clause (RC) interpretations. They found that subjects who had higher levels of exposure to the unexpected RC interpretation of the sentences had an easier time reading the RC sentences but a more difficult time reading the MV sentences. They concluded that syntactic adaptation occurs rapidly in unexpected structures and also results in difficulty with processing the previously expected alternative structures. This article presents two experiments. Experiment 1 was designed as a follow-up to Fine et al.'s study and failed to find evidence of adaptation. A power analysis of Fine et al.'s raw data revealed that a similar study would need double the items and four times the subjects to reach 95% power. In Experiment 2 we designed a close replication of Fine et al.'s experiment using these sample size guidelines. No evidence of rapid syntactic adaptation was found in this experiment. The failure to find evidence of adaptation in both experiments calls into question the robustness of the effect.

Under What Conditions Can Recursion Be Learned? Effects of Starting Small in Artificial Grammar Learning of Center-Embedded Structure

It has been suggested that external and/or internal limitations paradoxically may lead to superior learning, that is, the concepts of starting small and less is more (Elman, 1993; Newport, 1990). In this paper, we explore the type of incremental ordering during training that might help learning, and what mechanism explains this facilitation. We report four artificial grammar learning experiments with human participants. In Experiments 1a and 1b we found a beneficial effect of starting small using two types of simple recursive grammars: right‐branching and center‐embedding, with recursive embedded clauses in fixed positions and fixed length. This effect was replicated in Experiment 2 (N = 100). In Experiment 3 and 4, we used a more complex center‐embedded grammar with recursive loops in variable positions, producing strings of variable length. When participants were presented an incremental ordering of training stimuli, as in natural language, they were better able to generalize their knowledge of simple units to more complex units when the training input “grew” according to structural complexity, compared to when it “grew” according to string length. Overall, the results suggest that starting small confers an advantage for learning complex center‐embedded structures when the input is organized according to structural complexity.

Personalized learning: From neurogenetics of behaviors to designing optimal language training

Variability in drug responsivity has prompted the development of Personalized Medicine, which has shown great promise in utilizing genotypic information to develop safer and more effective drug regimens for patients. Similarly, individual variability in learning outcomes has puzzled researchers who seek to create optimal learning environments for students. "Personalized Learning" seeks to identify genetic, neural and behavioral predictors of individual differences in learning and aims to use predictors to help create optimal teaching paradigms. Evidence for Personalized Learning can be observed by connecting research in pharmacogenomics, cognitive genetics and behavioral experiments across domains of learning, which provides a framework for conducting empirical studies from the laboratory to the classroom and holds promise for addressing learning effectiveness in the individual learners. Evidence can also be seen in the subdomain of speech learning, thus providing initial support for the applicability of Personalized Learning to language.

Sampling Assumptions Affect Use of Indirect Negative Evidence in Language Learning

A classic debate in cognitive science revolves around understanding how children learn complex linguistic patterns, such as restrictions on verb alternations and contractions, without negative evidence. Recently, probabilistic models of language learning have been applied to this problem, framing it as a statistical inference from a random sample of sentences. These probabilistic models predict that learners should be sensitive to the way in which sentences are sampled. There are two main types of sampling assumptions that can operate in language learning: strong and weak sampling. Strong sampling, as assumed by probabilistic models, assumes the learning input is drawn from a distribution of grammatical samples from the underlying language and aims to learn this distribution. Thus, under strong sampling, the absence of a sentence construction from the input provides evidence that it has low or zero probability of grammaticality. Weak sampling does not make assumptions about the distribution from which the input is drawn, and thus the absence of a construction from the input as not used as evidence of its ungrammaticality. We demonstrate in a series of artificial language learning experiments that adults can produce behavior consistent with both sets of sampling assumptions, depending on how the learning problem is presented. These results suggest that people use information about the way in which linguistic input is sampled to guide their learning.

Discovery of a Recursive Principle: An Artificial Grammar Investigation of Human Learning of a Counting Recursion Language

Learning is typically understood as a process in which the behavior of an organism is progressively shaped until it closely approximates a target form. It is easy to comprehend how a motor skill or a vocabulary can be progressively learned-in each case, one can conceptualize a series of intermediate steps which lead to the formation of a proficient behavior. With grammar, it is more difficult to think in these terms. For example, center embedding recursive structures seem to involve a complex interplay between multiple symbolic rules which have to be in place simultaneously for the system to work at all, so it is not obvious how the mechanism could gradually come into being. Here, we offer empirical evidence from a new artificial language (or "artificial grammar") learning paradigm, Locus Prediction, that, despite the conceptual conundrum, recursion acquisition occurs gradually, at least for a simple formal language. In particular, we focus on a variant of the simplest recursive language, a (n) b (n) , and find evidence that (i) participants trained on two levels of structure (essentially ab and aabb) generalize to the next higher level (aaabbb) more readily than participants trained on one level of structure (ab) combined with a filler sentence; nevertheless, they do not generalize immediately; (ii) participants trained up to three levels (ab, aabb, aaabbb) generalize more readily to four levels than participants trained on two levels generalize to three; (iii) when we present the levels in succession, starting with the lower levels and including more and more of the higher levels, participants show evidence of transitioning between the levels gradually, exhibiting intermediate patterns of behavior on which they were not trained; (iv) the intermediate patterns of behavior are associated with perturbations of an attractor in the sense of dynamical systems theory. We argue that all of these behaviors indicate a theory of mental representation in which recursive systems lie on a continuum of grammar systems which are organized so that grammars that produce similar behaviors are near one another, and that people learning a recursive system are navigating progressively through the space of these grammars.

Regularization of languages by adults and children: A mathematical framework

The fascinating ability of humans to modify the linguistic input and "create" a language has been widely discussed. In the work of Newport and colleagues, it has been demonstrated that both children and adults have some ability to process inconsistent linguistic input and "improve" it by making it more consistent. In Hudson Kam and Newport (2009), artificial miniature language acquisition from an inconsistent source was studied. It was shown that (i) children are better at language regularization than adults and that (ii) adults can also regularize, depending on the structure of the input. In this paper we create a learning algorithm of the reinforcement-learning type, which exhibits patterns reported in Hudson Kam and Newport (2009) and suggests a way to explain them. It turns out that in order to capture the differences between children's and adults' learning patterns, we need to introduce a certain asymmetry in the learning algorithm. Namely, we have to assume that the reaction of the learners differs depending on whether or not the source's input coincides with the learner's internal hypothesis. We interpret this result in the context of a different reaction of children and adults to implicit, expectation-based evidence, positive or negative. We propose that a possible mechanism that contributes to the children's ability to regularize an inconsistent input is related to their heightened sensitivity to positive evidence rather than the (implicit) negative evidence. In our model, regularization comes naturally as a consequence of a stronger reaction of the children to evidence supporting their preferred hypothesis. In adults, their ability to adequately process implicit negative evidence prevents them from regularizing the inconsistent input, resulting in a weaker degree of regularization.

Transfer of conflict and cooperation from experienced games to new games: a connectionist model of learning

The question of whether, and if so how, learning can be transfered from previously experienced games to novel games has recently attracted the attention of the experimental game theory literature. Existing research presumes that learning operates over actions, beliefs or decision rules. This study instead uses a connectionist approach that learns a direct mapping from game payoffs to a probability distribution over own actions. Learning is operationalized as a backpropagation rule that adjusts the weights of feedforward neural networks in the direction of increasing the probability of an agent playing a myopic best response to the last game played. One advantage of this approach is that it expands the scope of the model to any possible n × n normal-form game allowing for a comprehensive model of transfer of learning. Agents are exposed to games drawn from one of seven classes of games with significantly different strategic characteristics and then forced to play games from previously unseen classes. I find significant transfer of learning, i.e., behavior that is path-dependent, or conditional on the previously seen games. Cooperation is more pronounced in new games when agents are previously exposed to games where the incentive to cooperate is stronger than the incentive to compete, i.e., when individual incentives are aligned. Prior exposure to Prisoner's dilemma, zero-sum and discoordination games led to a significant decrease in realized payoffs for all the game classes under investigation. A distinction is made between superficial and deep transfer of learning both—the former is driven by superficial payoff similarities between games, the latter by differences in the incentive structures or strategic implications of the games. I examine whether agents learn to play the Nash equilibria of games, how they select amongst multiple equilibria, and whether they transfer Nash equilibrium behavior to unseen games. Sufficient exposure to a strategically heterogeneous set of games is found to be a necessary condition for deep learning (and transfer) across game classes. Paradoxically, superficial transfer of learning is shown to lead to better outcomes than deep transfer for a wide range of game classes. The simulation results corroborate important experimental findings with human subjects, and make several novel predictions that can be tested experimentally.

The role of domain-general cognitive control in language comprehension

What role does domain-general cognitive control play in understanding linguistic input? Although much evidence has suggested that domain-general cognitive control and working memory resources are sometimes recruited during language comprehension, many aspects of this relationship remain elusive. For example, how frequently do cognitive control mechanisms get engaged when we understand language? And is this engagement necessary for successful comprehension? I here (a) review recent brain imaging evidence for the neural separability of the brain regions that support high-level linguistic processing vs. those that support domain-general cognitive control abilities; (b) define the space of possibilities for the relationship between these sets of brain regions; and (c) review the available evidence that constrains these possibilities to some extent. I argue that we should stop asking whether domain-general cognitive control mechanisms play a role in language comprehension, and instead focus on characterizing the division of labor between the cognitive control brain regions and the more functionally specialized language regions.

A matched filter hypothesis for cognitive control

The prefrontal cortex exerts top-down influences on several aspects of higher-order cognition by functioning as a filtering mechanism that biases bottom-up sensory information toward a response that is optimal in context. However, research also indicates that not all aspects of complex cognition benefit from prefrontal regulation. Here we review and synthesize this research with an emphasis on the domains of learning and creative cognition, and outline how the appropriate level of cognitive control in a given situation can vary depending on the organism's goals and the characteristics of the given task. We offer a matched filter hypothesis for cognitive control, which proposes that the optimal level of cognitive control is task-dependent, with high levels of cognitive control best suited to tasks that are explicit, rule-based, verbal or abstract, and can be accomplished given the capacity limits of working memory and with low levels of cognitive control best suited to tasks that are implicit, reward-based, non-verbal or intuitive, and which can be accomplished irrespective of working memory limitations. Our approach promotes a view of cognitive control as a tool adapted to a subset of common challenges, rather than an all-purpose optimization system suited to every problem the organism might encounter.

Language learning from positive evidence, reconsidered: a simplicity-based approach

Children learn their native language by exposure to their linguistic and communicative environment, but apparently without requiring that their mistakes be corrected. Such learning from "positive evidence" has been viewed as raising "logical" problems for language acquisition. In particular, without correction, how is the child to recover from conjecturing an over-general grammar, which will be consistent with any sentence that the child hears? There have been many proposals concerning how this "logical problem" can be dissolved. In this study, we review recent formal results showing that the learner has sufficient data to learn successfully from positive evidence, if it favors the simplest encoding of the linguistic input. Results include the learnability of linguistic prediction, grammaticality judgments, language production, and form-meaning mappings. The simplicity approach can also be "scaled down" to analyze the learnability of specific linguistic constructions, and it is amenable to empirical testing as a framework for describing human language acquisition.

How semantic biases in simple adjacencies affect learning a complex structure with non-adjacencies in AGL: a statistical account

A major theoretical debate in language acquisition research regards the learnability of hierarchical structures. The artificial grammar learning methodology is increasingly influential in approaching this question. Studies using an artificial centre-embedded A(n)B(n) grammar without semantics draw conflicting conclusions. This study investigates the facilitating effect of distributional biases in simple AB adjacencies in the input sample--caused in natural languages, among others, by semantic biases-on learning a centre-embedded structure. A mathematical simulation of the linguistic input and the learning, comparing various distributional biases in AB pairs, suggests that strong distributional biases might help us to grasp the complex A(n)B(n) hierarchical structure in a later stage. This theoretical investigation might contribute to our understanding of how distributional features of the input--including those caused by semantic variation--help learning complex structures in natural languages.

The role of prior experience in language acquisition

Learners exposed to an artificial language recognize its abstract structural regularities when instantiated in a novel vocabulary (e.g., Gómez, Gerken, & Schvaneveldt, 2000; Tunney & Altmann, 2001). We asked whether such sensitivity accelerates subsequent learning, and enables acquisition of more complex structure. In Experiment 1, pre-exposure to a category-induction language of the form aX bY sped subsequent learning when the language is instantiated in a different vocabulary. In Experiment 2, while naíve learners did not acquire an acX bcY language, in which aX and bY co-occurrence regularities were separated by a c-element, prior experience with an aX bY language provided some benefit. In Experiment 3 we replicated this finding with a 24-hour delay between learning phases, and controlled for prior experience with the aX bY language's prosodic and phonological characteristics. These findings suggest that learners, and the structure they can acquire, change as a function of experience.

Increasing the odds: applying emergentist theory in language intervention

PURPOSE

This review introduces emergentism, which is a leading theory of language development that states that language ability is the product of interactions between the child's language environment and his or her learning capabilities. The review suggests ways in which emergentism provides a theoretical rationale for interventions that are designed to address developmental language delays in young children.

METHOD

A review of selected literature on emergentist theory and research is presented, with a focus on the acquisition of early morphology and syntax. A significant method for developing and testing emergentist theory, connectionist modeling, is described. Key themes from both connectionist and behavioral studies are summarized and applied with specific examples to language intervention techniques. A case study is presented to integrate elements of emergentism with language intervention.

CONCLUSIONS

Evaluating the theoretical foundation for language interventions is an important step in evidence-based practice. This article introduces three themes in the emergentist literature that have implications for language intervention: (a) sufficiency of language input, (b) active engagement of the child with the input, and (c) factors that increase the odds for correctly mapping language form to meaning. Evidence supporting the importance of these factors in effective language intervention is presented, along with limitations in that evidence.