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

We investigated whether readers use the low-level cue of proper noun capitalization in the parafovea to infer syntactic category, and whether this results in an early update of the representation of a sentence’s syntactic structure. Participants read sentences containing either a subject relative or object relative clause, in which the relative clause’s overt argument was a proper noun (e.g., The tall lanky guard who alerted Charlie/Charlie alerted to the danger was young) across three experiments. In Experiment 1 these sentences were presented in normal sentence casing or entirely in upper case. In Experiment 2 participants received either valid or invalid parafoveal previews of the relative clause. In Experiment 3 participants viewed relative clauses in only normal conditions. We hypothesized that we would observe relative clause effects (i.e., inflated fixation times for object relative clauses) while readers were still fixated on the word who, if readers use capitalization to infer a parafoveal word’s syntactic class. This would constitute a syntactic parafoveal-on-foveal effect. Furthermore, we hypothesised that this effect should be influenced by sentence casing in Experiment 1 (with no cue for syntactic category being available in upper case sentences) but not by parafoveal preview validity of the target words. We observed syntactic parafoveal-on-foveal effects in Experiment 1 and 3, and a Bayesian analysis of the combined data from all three experiments. These effects seemed to be influenced more by noun capitalization than lexical processing. We discuss our findings in relation to models of eye movement control and sentence processing theories.

Abstract

In two eye-tracking studies we investigated whether readers can detect a violation of the phonological-grammatical convention for the indefinite article an to be followed by a word beginning with a vowel when these two words appear in the parafovea. Across two experiments participants read sentences in which the word an was followed by a parafoveal preview that was either correct (e.g. Icelandic), incorrect and represented a phonological violation (e.g. Mongolian), or incorrect without representing a phonological violation (e.g. Ethiopian), with this parafoveal preview changing to the target word as participants made a saccade into the space preceding an. Our data suggests that participants detected the phonological violation while the target word was still two words to the right of fixation, with participants making more regressions from the previewed word and having longer go-past times on this word when they received a violation preview as opposed to a non-violation preview. We argue that participants were attempting to perform aspects of sentence integration on the basis of low-level orthographic information from the previewed word.

Abstract

We present an eye-tracking study testing a hypothesis emerging from several theories of prediction during language processing, whereby predictable words should be skipped more than unpredictable words even in syntactically illegal positions. Participants read sentences in which a target word became predictable by a certain point (e.g., “bone” is 92% predictable given, “The dog buried his. . .”), with the next word actually being an intensifier (e.g., “really”), which a noun cannot follow. The target noun remained predictable to appear later in the sentence. We used the boundary paradigm to present the predictable noun or an alternative unpredictable noun (e.g., “food”) directly after the intensifier, until participants moved beyond the intensifier, at which point the noun changed to a syntactically legal word. Participants also read sentences in which predictable or unpredictable nouns appeared in syntactically legal positions. A Bayesian linear-mixed model suggested a 5.7% predictability effect on skipping of nouns in syntactically legal positions, and a 3.1% predictability effect on skipping of nouns in illegal positions. We discuss our findings in relation to theories of lexical prediction during reading.

Abstract

Humans readily generalize, applying prior knowledge to novel situations and stimuli. Advances in machine learning have begun to approximate and even surpass human performance, but these systems struggle to generalize what they have learned to untrained situations. We present a model based on wellestablished neurocomputational principles that demonstrates human-level generalisation. This model is trained to play one video game (Breakout) and performs one-shot generalisation to a new game (Pong) with different characteristics. The model
generalizes because it learns structured representations that are functionally symbolic (viz., a role-filler binding calculus) from unstructured training data. It does so without feedback, and without requiring that structured representations are specified a priori. Specifically, the model uses neural co-activation to discover which characteristics of the input are invariant and to learn relational predicates, and oscillatory regularities in network firing to bind predicates to arguments. To our knowledge,
this is the first demonstration of human-like generalisation in a machine system that does not assume structured representa-
tions to begin with.

Abstract

The representations generated by many mod-
els of language (word embeddings, recurrent
neural networks and transformers) correlate
to brain activity recorded while people read.
However, these decoding results are usually
based on the brain’s reaction to syntactically
and semantically sound language stimuli. In
this study, we asked: how does an LSTM (long
short term memory) language model, trained
(by and large) on semantically and syntac-
tically intact language, represent a language
sample with degraded semantic or syntactic
information? Does the LSTM representation
still resemble the brain’s reaction? We found
that, even for some kinds of nonsensical lan-
guage, there is a statistically significant rela-
tionship between the brain’s activity and the
representations of an LSTM. This indicates
that, at least in some instances, LSTMs and the
human brain handle nonsensical data similarly.

Abstract

Understanding spoken language requires the integration and weighting of multiple cues, and may call on cue integration mechanisms that have been studied in other areas of perception. In the current study, we used eye-tracking (visual-world paradigm) to examine how contextual speech rate (a lower-level, perceptual cue) and morphosyntactic knowledge (a higher-level, linguistic cue) are iteratively combined and integrated. Results indicate that participants used contextual rate information immediately, which we interpret as evidence of perceptual inference and the generation of predictions about upcoming morphosyntactic information. Additionally, we observed that early rate effects remained active in the presence of later conflicting lexical information. This result demonstrates that (1) contextual speech rate functions as a cue to morphosyntactic inferences, even in the presence of subsequent disambiguating information; and (2) listeners iteratively use multiple sources of information to draw inferences and generate predictions during speech comprehension. We discuss the implication of these demonstrations for theories of language processing

Abstract

During spoken language comprehension, listeners make use of both knowledge-based and signal-based sources of information, but little is known about how cues from these distinct levels of representational hierarchy are weighted and integrated online. In an eye-tracking experiment using the visual world paradigm, we investigated the flexible weighting and integration of morphosyntactic gender marking (a knowledge-based cue) and contextual speech rate (a signal-based cue). We observed that participants used the morphosyntactic cue immediately to make predictions about upcoming referents, even in the presence of uncertainty about the cue’s reliability. Moreover, we found speech rate normalization effects in participants’ gaze patterns even in the presence of preceding morphosyntactic information. These results demonstrate that cues are weighted and integrated flexibly online, rather than adhering to a strict hierarchy. We further found rate normalization effects in the looking behavior of participants who showed a strong behavioral preference for the morphosyntactic gender cue. This indicates that rate normalization effects are robust and potentially automatic. We discuss these results in light of theories of cue integration and the two-stage model of acoustic context effects

Abstract

Neural oscillations track linguistic information during speech comprehension (e.g., Ding et al., 2016; Keitel et al., 2018), and are known to be modulated by acoustic landmarks and speech intelligibility (e.g., Doelling et al., 2014; Zoefel & VanRullen, 2015). However, studies investigating linguistic tracking have either relied on non-naturalistic isochronous stimuli or failed to fully control for prosody. Therefore, it is still unclear whether low frequency activity tracks linguistic structure during natural speech, where linguistic structure does not follow such a palpable temporal pattern. Here, we measured electroencephalography (EEG) and manipulated the presence of semantic and syntactic information apart from the timescale of their occurrence, while carefully controlling for the acoustic-prosodic and lexical-semantic information in the signal. EEG was recorded while 29 adult native speakers (22 women, 7 men) listened to naturally-spoken Dutch sentences, jabberwocky controls with morphemes and sentential prosody, word lists with lexical content but no phrase structure, and backwards acoustically-matched controls. Mutual information (MI) analysis revealed sensitivity to linguistic content: MI was highest for sentences at the phrasal (0.8-1.1 Hz) and lexical timescale (1.9-2.8 Hz), suggesting that the delta-band is modulated by lexically-driven combinatorial processing beyond prosody, and that linguistic content (i.e., structure and meaning) organizes neural oscillations beyond the timescale and rhythmicity of the stimulus. This pattern is consistent with neurophysiologically inspired models of language comprehension (Martin, 2016, 2020; Martin & Doumas, 2017) where oscillations encode endogenously generated linguistic content over and above exogenous or stimulus-driven timing and rhythm information.

Abstract

Hierarchical structure and compositionality imbue human language with unparalleled expressive power and set it apart from other perception–action systems. However, neither formal nor neurobiological models account for how these defining computational properties might arise in a physiological system. I attempt to reconcile hierarchy and compositionality with principles from cell assembly computation in neuroscience; the result is an emerging theory of how the brain could convert distributed perceptual representations into hierarchical structures across multiple timescales while representing interpretable incremental stages of (de) compositional meaning. The model's architecture—a multidimensional coordinate system based on neurophysiological models of sensory processing—proposes that a manifold of neural trajectories encodes sensory, motor, and abstract linguistic states. Gain modulation, including inhibition, tunes the path in the manifold in accordance with behavior and is how latent structure is inferred. As a consequence, predictive information about upcoming sensory input during production and comprehension is available without a separate operation. The proposed processing mechanism is synthesized from current models of neural entrainment to speech, concepts from systems neuroscience and category theory, and a symbolic-connectionist computational model that uses time and rhythm to structure information. I build on evidence from cognitive neuroscience and computational modeling that suggests a formal and mechanistic alignment between structure building and neural oscillations and moves toward unifying basic insights from linguistics and psycholinguistics with the currency of neural computation.

Abstract

Research on speech processing is often focused on a phenomenon termed “entrainment”, whereby the cortex shadows rhythmic acoustic information with oscillatory activity. Entrainment has been observed to a range of rhythms present in speech; in addition, synchronicity with abstract information (e.g. syntactic structures) has been observed. Entrainment accounts face two challenges: First, speech is not exactly rhythmic; second, synchronicity with representations that lack a clear acoustic counterpart has been described. We propose that apparent entrainment does not always result from acoustic information. Rather, internal rhythms may have functionalities in the generation of abstract representations and predictions. While acoustics may often provide punctate opportunities for entrainment, internal rhythms may also live a life of their own to infer and predict information, leading to intrinsic synchronicity – not to be counted as entrainment. This possibility may open up new research avenues in the psycho– and neurolinguistic study of language processing and language development.

Abstract

Could meaning be read from acoustics, or from the refraction rate of pyramidal cells innervated by the cochlea, everyone would be an omniglot. Speech does not contain sufficient acoustic cues to identify linguistic units such as morphemes, words, and phrases without prior knowledge. Our target article (Meyer, L., Sun, Y., & Martin, A. E. (2019). Synchronous, but not entrained: Exogenous and endogenous cortical rhythms of speech and language processing. Language, Cognition and Neuroscience, 1–11. https://doi.org/10.1080/23273798.2019.1693050) thus questioned the concept of “entrainment” of neural oscillations to such units. We suggested that synchronicity with these points to the existence of endogenous functional “oscillators”—or population rhythmic activity in Giraud’s (2020) terms—that underlie the inference, generation, and prediction of linguistic units. Here, we address a series of inspirational commentaries by our colleagues. As apparent from these, some issues raised by our target article have already been raised in the literature. Psycho– and neurolinguists might still benefit from our reply, as “oscillations are an old concept in vision and motor functions, but a new one in linguistics” (Giraud, A.-L. 2020. Oscillations for all A commentary on Meyer, Sun & Martin (2020). Language, Cognition and Neuroscience, 1–8).

Abstract

How a system represents information tightly constrains the kinds of problems it can solve. Humans routinely solve problems that appear to require structured representations of stimulus properties and the relations between them. An account of how we might acquire such representations has central importance for theories of human cognition. We describe how a system can learn structured relational representations from initially unstructured inputs using comparison, sensitivity to time, and a modified Hebbian learning algorithm. We summarize how the model DORA (Discovery of Relations by Analogy) instantiates this approach, which we call predicate learning, as well as how the model captures several phenomena from cognitive development, relational reasoning, and language processing in the human brain. Predicate learning offers a link between models based on formal languages and models which learn from experience and provides an existence proof for how structured representations might be learned in the first place.

Abstract

In response to recommendations to redefine statistical significance to P ≤ 0.005, we propose that researchers should transparently report and justify all choices they make when designing a study, including the alpha level.

Abstract

Language processing requires us to integrate incoming linguistic representations with representations of past input, often across intervening words and phrases. This computational situation has been argued to require retrieval of the appropriate representations from memory via a set of features or representations serving as retrieval cues. However, even within in a cue-based retrieval account of language comprehension, both the structure of retrieval cues and the particular computation that underlies direct-access retrieval are still underspecified. Evidence from two event-related brain potential (ERP) experiments that show cue-based interference from different types of linguistic representations during ellipsis comprehension are consistent with an architecture wherein different cue types are integrated, and where the interaction of cue with the recent contents of memory determines processing outcome, including expression of the interference effect in ERP componentry. I conclude that retrieval likely includes a computation where cues are integrated with the contents of memory via a linear weighting scheme, and I propose vector addition as a candidate formalization of this computation. I attempt to account for these effects and other related phenomena within a broader cue-based framework of language processing.

Abstract

Language comprehension involves coping with ambiguity and recovering from misanalysis. Syntactic ambiguity resolution is associated with increased reading times, a classic finding that has shaped theories of sentence processing. However, reaction times conflate the time it takes a process to complete with the quality of the behavior-related information available to the system. We therefore used the speed-accuracy tradeoff procedure (SAT) to derive orthogonal estimates of processing time and interpretation accuracy, and tested whether stronger retrieval cues (via semantic relatedness: neighed->horse vs. fell->horse) aid interpretation during recovery. On average, ambiguous sentences took 250ms longer (SAT rate) to interpret than unambiguous controls, demonstrating veridical differences in processing time. Retrieval cues more strongly related to the true subject always increased accuracy, regardless of ambiguity. These findings are consistent with a language processing architecture where cue-driven operations give rise to interpretation, and wherein diagnostic cues aid retrieval, regardless of parsing difficulty or structural uncertainty.

Abstract

Two experiments examined the nature of the phonological representations used during visual word recognition. We tested whether a minimality constraint (R. Frost, 1998) limits the complexity of early representations to a simple string of phonemes. Alternatively, readers might activate elaborated representations that include prosodic syllable information before lexical access. In a modified lexical decision task (Experiment 1), words were preceded by parafoveal previews that were congruent with a target's initial syllable as well as previews that contained 1 letter more or less than the initial syllable. Lexical decision times were faster in the syllable congruent conditions than in the incongruent conditions. In Experiment 2, we recorded brain electrical potentials (electroencephalograms) during single word reading in a masked priming paradigm. The event-related potential waveform elicited in the syllable congruent condition was more positive 250-350 ms posttarget compared with the waveform elicited in the syllable incongruent condition. In combination, these experiments demonstrate that readers process prosodic syllable information early in visual word recognition in English. They offer further evidence that skilled readers routinely activate elaborated, speechlike phonological representations during silent reading. (PsycINFO Database Record (c) 2016 APA, all rights reserved)

Abstract

Interpreting a verb-phrase ellipsis (VP ellipsis) requires accessing an antecedent in memory, and then integrating a representation of this antecedent into the local context. We investigated the online interpretation of VP ellipsis in an eye-tracking experiment and four speed–accuracy tradeoff experiments. To investigate whether the antecedent for a VP ellipsis is accessed with a search or direct-access retrieval process, Experiments 1 and 2 measured the effect of the distance between an ellipsis and its antecedent on the speed and accuracy of comprehension. Accuracy was lower with longer distances, indicating that interpolated material reduced the quality of retrieved information about the antecedent. However, contra a search process, distance did not affect the speed of interpreting ellipsis. This pattern suggests that antecedent representations are content-addressable and retrieved with a direct-access process. To determine whether interpreting ellipsis involves copying antecedent information into the ellipsis site, Experiments 3–5 manipulated the length and complexity of the antecedent. Some types of antecedent complexity lowered accuracy, notably, the number of discourse entities in the antecedent. However, neither antecedent length nor complexity affected the speed of interpreting the ellipsis. This pattern is inconsistent with a copy operation, and it suggests that ellipsis interpretation may involve a pointer to extant structures in memory.