Publications

Abstract

In this study, we compared the performance of a long short-term memory (LSTM) neural network to the behavior of human participants on a language task that requires hierarchically structured knowledge. We show that humans interpret ambiguous noun phrases, such as second blue ball, in line with their hierarchical constituent structure. LSTMs, instead, only do
so after unambiguous training, and they do not systematically generalize to novel items. Overall, the results of our simulations indicate that a model can behave hierarchically without relying on hierarchical constituent structure.

Abstract

How a system represents information tightly constrains the kinds of problems it can solve. Humans routinely solve problems that appear to require abstract representations of stimulus properties and relations. How we acquire such representations has central importance in an account of human cognition. We briefly describe a theory of how a system can learn invariant responses to instances of similarity and relative magnitude, and how structured, relational representations can be learned from initially unstructured inputs. Two operations, comparing distributed representations and learning from the concomitant network dynamics in time, underpin the ability to learn these representations and to respond to invariance in the environment. Comparing analog representations of absolute magnitude produces invariant signals that carry information about similarity and relative magnitude. We describe how a system can then use this information to bootstrap learning structured (i.e., symbolic) concepts of relative magnitude from experience without assuming such representations a priori.

Abstract

Psychology endeavors to develop theories of human capacities and behaviors on the basis of a variety of methodologies and dependent measures. We argue that one of the most divisive factors in psychological science is whether researchers choose to use computational modeling of theories (over and above data) during the scientific-inference process. Modeling is undervalued yet holds promise for advancing psychological science. The inherent demands of computational modeling guide us toward better science by forcing us to conceptually analyze, specify, and formalize intuitions that otherwise remain unexamined—what we dub open theory. Constraining our inference process through modeling enables us to build explanatory and predictive theories. Here, we present scientific inference in psychology as a path function in which each step shapes the next. Computational modeling can constrain these steps, thus advancing scientific inference over and above the stewardship of experimental practice (e.g., preregistration). If psychology continues to eschew computational modeling, we predict more replicability crises and persistent failure at coherent theory building. This is because without formal modeling we lack open and transparent theorizing. We also explain how to formalize, specify, and implement a computational model, emphasizing that the advantages of modeling can be achieved by anyone with benefit to all.

Abstract

Whether neural networks can capture relational knowledge is a matter of long-standing controversy. Recently, some researchers have argued that (1) classic connectionist models can handle relational structure and (2) the success of deep learning approaches to natural language processing suggests that structured representations are unnecessary to model human language. We tested the Story Gestalt model, a classic connectionist model of text comprehension, and a Sequence-to-Sequence with Attention model, a modern deep learning architecture for natural language processing. Both models were trained to answer questions about stories based on abstract thematic roles. Two simulations varied the statistical structure of new stories while keeping their relational structure intact. The performance of each model fell below chance at least under one manipulation. We argue that both models fail our tests because they can't perform dynamic binding. These results cast doubts on the suitability of traditional neural networks for explaining relational reasoning and language processing phenomena.

Abstract

Neuronal oscillations putatively track speech in order to optimize sensory processing. However, it is unclear how isochronous brain oscillations can track pseudo-rhythmic speech input. Here we propose that oscillations can track pseudo-rhythmic speech when considering that speech time is dependent on content-based predictions flowing from internal language models. We show that temporal dynamics of speech are dependent on the predictability of words in a sentence. A computational model including oscillations, feedback, and inhibition is able to track pseudo-rhythmic speech input. As the model processes, it generates temporal phase codes, which are a candidate mechanism for carrying information forward in time. The model is optimally sensitive to the natural temporal speech dynamics and can explain empirical data on temporal speech illusions. Our results suggest that speech tracking does not have to rely only on the acoustics but could also exploit ongoing interactions between oscillations and constraints flowing from internal language models.

Abstract

Computation in neuronal assemblies is putatively reflected in the excitatory and inhibitory cycles of activation distributed throughout the brain. In speech and language processing, coordination of these cycles resulting in phase synchronization has been argued to reflect the integration of information on different timescales (e.g. segmenting acoustics signals to phonemic and syllabic representations; (Giraud and Poeppel 2012 Nat. Neurosci.15, 511 (doi:10.1038/nn.3063)). A natural extension of this claim is that phase synchronization functions similarly to support the inference of more abstract higher-level linguistic structures (Martin 2016 Front. Psychol.7, 120; Martin and Doumas 2017 PLoS Biol. 15, e2000663 (doi:10.1371/journal.pbio.2000663); Martin and Doumas. 2019 Curr. Opin. Behav. Sci.29, 77–83 (doi:10.1016/j.cobeha.2019.04.008)). Hale et al. (Hale et al. 2018 Finding syntax in human encephalography with beam search. arXiv 1806.04127 (http://arxiv.org/abs/1806.04127)) showed that syntactically driven parsing decisions predict electroencephalography (EEG) responses in the time domain; here we ask whether phase synchronization in the form of either inter-trial phrase coherence or cross-frequency coupling (CFC) between high-frequency (i.e. gamma) bursts and lower-frequency carrier signals (i.e. delta, theta), changes as the linguistic structures of compositional meaning (viz., bracket completions, as denoted by the onset of words that complete phrases) accrue. We use a naturalistic story-listening EEG dataset from Hale et al. to assess the relationship between linguistic structure and phase alignment. We observe increased phase synchronization as a function of phrase counts in the delta, theta, and gamma bands, especially for function words. A more complex pattern emerged for CFC as phrase count changed, possibly related to the lack of a one-to-one mapping between ‘size’ of linguistic structure and frequency band—an assumption that is tacit in recent frameworks. These results emphasize the important role that phase synchronization, desynchronization, and thus, inhibition, play in the construction of compositional meaning by distributed neural networks in the brain.

Abstract

Human thought and language have extraordinary expressive power because meaningful parts can be assembled into more complex semantic structures. This partly underlies our ability to compose meanings into endlessly novel configurations, and sets us apart from other species and current computing devices. Crucially, human behaviour, including language use and linguistic data, indicates that composing parts into complex structures does not threaten the existence of constituent parts as independent units in the system: parts and wholes exist simultaneously yet independently from one another in the mind and brain. This independence is evident in human behaviour, but it seems at odds with what is known about the brain's exquisite sensitivity to statistical patterns: everyday language use is productive and expressive precisely because it can go beyond statistical regularities. Formal theories in philosophy and linguistics explain this fact by assuming that language and thought are compositional: systems of representations that separate a variable (or role) from its values (fillers), such that the meaning of a complex expression is a function of the values assigned to the variables. The debate on whether and how compositional systems could be implemented in minds, brains and machines remains vigorous. However, it has not yet resulted in mechanistic models of semantic composition: how, then, are the constituents of thoughts and sentences put and held together? We review and discuss current efforts at understanding this problem, and we chart possible routes for future research.

Abstract

Neither neurobiological nor process models of meaning composition specify the operator through which constituent parts are bound together into compositional structures. In this paper, we argue that a neurophysiological computation system cannot achieve the compositionality exhibited in human thought and language if it were to rely on a multiplicative operator to perform binding, as the tensor product (TP)-based systems that have been widely adopted in cognitive science, neuroscience and artificial intelligence do. We show via simulation and two behavioural experiments that TPs violate variable-value independence, but human behaviour does not. Specifically, TPs fail to capture that in the statements fuzzy cactus and fuzzy penguin, both cactus and penguin are predicated by fuzzy(x) and belong to the set of fuzzy things, rendering these arguments similar to each other. Consistent with that thesis, people judged arguments that shared the same role to be similar, even when those arguments themselves (e.g., cacti and penguins) were judged to be dissimilar when in isolation. By contrast, the similarity of the TPs representing fuzzy(cactus) and fuzzy(penguin) was determined by the similarity of the arguments, which in this case approaches zero. Based on these results, we argue that neural systems that use TPs for binding cannot approximate how the human mind and brain represent compositional information during processing. We describe a contrasting binding mechanism that any physiological or artificial neural system could use to maintain independence between a role and its argument, a prerequisite for compositionality and, thus, for instantiating the expressive power of human thought and language in a neural system.

Abstract

Humans learn to represent complex structures (e.g. natural language, music, mathematics) from experience with their environments. Often such structures are latent, hidden, or not encoded in statistics about sensory representations alone. Accounts of human cognition have long emphasized the importance of structured representations, yet the majority of contemporary neural networks do not learn structure from experience. Here, we describe one way that structured, functionally symbolic representations can be instantiated in an artificial neural network. Then, we describe how such latent structures (viz. predicates) can be learned from experience with unstructured data. Our approach exploits two principles from psychology and neuroscience: comparison of representations, and the naturally occurring dynamic properties of distributed computing across neuronal assemblies (viz. neural oscillations). We discuss how the ability to learn predicates from experience, to represent information compositionally, and to extrapolate knowledge to unseen data is core to understanding and modeling the most complex human behaviors (e.g. relational reasoning, analogy, language processing, game play).

Abstract

Abstract mental representation is fundamental for human cognition. Forming such representations in time, especially from dynamic and noisy perceptual input, is a challenge for any processing modality, but perhaps none so acutely as for language processing. We show that LISA (Hummel & Holyaok, 1997) and DORA (Doumas, Hummel, & Sandhofer, 2008), models built to process and to learn structured (i.e., symbolic) rep resentations of conceptual properties and relations from unstructured inputs, show oscillatory activation during processing that is highly similar to the cortical activity elicited by the linguistic stimuli from Ding et al.(2016). We argue, as Ding et al.(2016), that this activation reflects formation of hierarchical linguistic representation, and furthermore, that the kind of computational mechanisms in LISA/DORA (e.g., temporal binding by systematic asynchrony of firing) may underlie formation of abstract linguistic representations in the human brain. It may be this repurposing that allowed for the generation or mergence of hierarchical linguistic structure, and therefore, human language, from extant cognitive and neural systems. We conclude that models of thinking and reasoning and models of language processing must be integrated —not only for increased plausiblity, but in order to advance both fields towards a larger integrative model of human cognition

Abstract

We used ERPs to investigate the pre-activation of form and meaning in language comprehension. Participants read high-cloze sentence contexts (e.g., “The student is going to the library to borrow a…”), followed by a word that was predictable (book), form-related (hook) or semantically related (page) to the predictable word, or unrelated (sofa). At a 500 ms SOA (Experiment 1), semantically related words, but not form-related words, elicited a reduced N400 compared to unrelated words. At a 700 ms SOA (Experiment 2), semantically related words and form-related words elicited reduced N400 effects, but the effect for form-related words occurred in very high-cloze sentences only. At both SOAs, form-related words elicited an enhanced, post-N400 posterior positivity (Late Positive Component effect). The N400 effects suggest that readers can pre-activate meaning and form information for highly predictable words, but form pre-activation is more limited than meaning pre-activation. The post-N400 LPC effect suggests that participants detected the form similarity between expected and encountered input. Pre-activation of word forms crucially depends upon the time that readers have to make predictions, in line with production-based accounts of linguistic prediction.

Abstract

I argue that cue integration, a psychophysiological mechanism from vision and multisensory perception, offers a computational linking hypothesis between psycholinguistic theory and neurobiological models of language. I propose that this mechanism, which incorporates probabilistic estimates of a cue's reliability, might function in language processing from the perception of a phoneme to the comprehension of a phrase structure. I briefly consider the implications of the cue integration hypothesis for an integrated theory of language that includes acquisition, production, dialogue and bilingualism, while grounding the hypothesis in canonical neural computation.

Abstract

Language comprehension requires recovering meaning from linguistic form, even when the mapping between the two is indirect. A canonical example is ellipsis, the omission of information that is subsequently understood without being overtly pronounced. Comprehension of ellipsis requires retrieval of an antecedent from memory, without prior prediction, a property which enables the study of retrieval in situ ( Martin and McElree, 2008 and Martin and McElree, 2009). Sluicing, or inflectional-phrase ellipsis, in the presence of a conjunction, presents a test case where a competing antecedent position is syntactically licensed, in contrast with most cases of nonadjacent dependency, including verb–phrase ellipsis. We present speed–accuracy tradeoff and eye-movement data inconsistent with the hypothesis that retrieval is accomplished via a syntactically guided search, a particular variant of search not examined in past research. The observed timecourse profiles are consistent with the hypothesis that antecedents are retrieved via a cue-dependent direct-access mechanism susceptible to general memory variables.

Abstract

Comprehension of verb-phrase ellipsis (VPE) requires reevaluation of recently processed constituents, which often necessitates retrieval of information about the elided constituent from memory. A. E. Martin and B. McElree (2008) argued that representations formed during comprehension are content addressable and that VPE antecedents are retrieved from memory via a cue-dependent direct-access pointer rather than via a search process. This hypothesis was further tested by manipulating the location of interfering material—either before the onset of the antecedent (proactive interference; PI) or intervening between antecedent and ellipsis site (retroactive interference; RI). The speed–accuracy tradeoff procedure was used to measure the time course of VPE processing. The location of the interfering material affected VPE comprehension accuracy: RI conditions engendered lower accuracy than PI conditions. Crucially, location did not affect the speed of processing VPE, which is inconsistent with both forward and backward search mechanisms. The observed time-course profiles are consistent with the hypothesis that VPE antecedents are retrieved via a cue-dependent direct-access operation. (PsycINFO Database Record (c) 2016 APA, all rights reserved)

Abstract

To study the neural bases of semantic composition in language processing without confounds from syntactic composition, recent magnetoencephalography (MEG) studies have investigated the processing of constructions that exhibit some type of syntax-semantics mismatch. The most studied case of such a mismatch is complement coercion; expressions such as the author began the book, where an entity-denoting noun phrase is coerced into an eventive meaning in order to match the semantic properties of the event-selecting verb (e.g., ‘the author began reading/writing the book’). These expressions have been found to elicit increased activity in the Anterior Midline Field (AMF), an MEG component elicited at frontomedial sensors at ∼400 ms after the onset of the coercing noun [Pylkkänen, L., & McElree, B. (2007). An MEG study of silent meaning. Journal of Cognitive Neuroscience, 19, 11]. Thus, the AMF constitutes a potential neural correlate of coercion. However, the AMF was generated in ventromedial prefrontal regions, which are heavily associated with decision-making. This raises the possibility that, instead of semantic processing, the AMF effect may have been related to the experimental task, which was a sensicality judgment. We tested this hypothesis by assessing the effect of coercion when subjects were simply reading for comprehension, without a decision-task. Additionally, we investigated coercion in an adjectival rather than a verbal environment to further generalize the findings. Our results show that an AMF effect of coercion is elicited without a decision-task and that the effect also extends to this novel syntactic environment. We conclude that in addition to its role in non-linguistic higher cognition, ventromedial prefrontal regions contribute to the resolution of syntax-semantics mismatches in language processing.