Neurobehavioral Correlates of Surprisal in Language Comprehension: A Neurocomputational Model

Brain and grammar: revealing electrophysiological basic structures with competing statistical models

Acoustic, lexical, and syntactic information are simultaneously processed in the brain requiring complex strategies to distinguish their electrophysiological activity. Capitalizing on previous works that factor out acoustic information, we could concentrate on the lexical and syntactic contribution to language processing by testing competing statistical models. We exploited electroencephalographic recordings and compared different surprisal models selectively involving lexical information, part of speech, or syntactic structures in various combinations. Electroencephalographic responses were recorded in 32 participants during listening to affirmative active declarative sentences. We compared the activation corresponding to basic syntactic structures, such as noun phrases vs. verb phrases. Lexical and syntactic processing activates different frequency bands, partially different time windows, and different networks. Moreover, surprisal models based on part of speech inventory only do not explain well the electrophysiological data, while those including syntactic information do. By disentangling acoustic, lexical, and syntactic information, we demonstrated differential brain sensitivity to syntactic information. These results confirm and extend previous measures obtained with intracranial recordings, supporting our hypothesis that syntactic structures are crucial in neural language processing. This study provides a detailed understanding of how the brain processes syntactic information, highlighting the importance of syntactic surprisal in shaping neural responses during language comprehension.

On the Mathematical Relationship Between Contextual Probability and N400 Amplitude

Accounts of human language comprehension propose different mathematical relationships between the contextual probability of a word and how difficult it is to process, including linear, logarithmic, and super-logarithmic ones. However, the empirical evidence favoring any of these over the others is mixed, appearing to vary depending on the index of processing difficulty used and the approach taken to calculate contextual probability. To help disentangle these results, we focus on the mathematical relationship between corpus-derived contextual probability and the N400, a neural index of processing difficulty. Specifically, we use 37 contemporary transformer language models to calculate the contextual probability of stimuli from 6 experimental studies of the N400, and test whether N400 amplitude is best predicted by a linear, logarithmic, super-logarithmic, or sub-logarithmic transformation of the probabilities calculated using these language models, as well as combinations of these transformed metrics. We replicate the finding that on some datasets, a combination of linearly and logarithmically-transformed probability can predict N400 amplitude better than either metric alone. In addition, we find that overall, the best single predictor of N400 amplitude is sub-logarithmically-transformed probability, which for almost all language models and datasets explains all the variance in N400 amplitude otherwise explained by the linear and logarithmic transformations. This is a novel finding that is not predicted by any current theoretical accounts, and thus one that we argue is likely to play an important role in increasing our understanding of how the statistical regularities of language impact language comprehension.

Tracking Lexical and Semantic Prediction Error Underlying the N400 Using Artificial Neural Network Models of Sentence Processing

Recent research has shown that the internal dynamics of an artificial neural network model of sentence comprehension displayed a similar pattern to the amplitude of the N400 in several conditions known to modulate this event-related potential. These results led Rabovsky et al. (2018) to suggest that the N400 might reflect change in an implicit predictive representation of meaning corresponding to semantic prediction error. This explanation stands as an alternative to the hypothesis that the N400 reflects lexical prediction error as estimated by word surprisal (Frank et al., 2015). In the present study, we directly model the amplitude of the N400 elicited during naturalistic sentence processing by using as predictor the update of the distributed representation of sentence meaning generated by a sentence gestalt model (McClelland et al., 1989) trained on a large-scale text corpus. This enables a quantitative prediction of N400 amplitudes based on a cognitively motivated model, as well as quantitative comparison of this model to alternative models of the N400. Specifically, we compare the update measure from the sentence gestalt model to surprisal estimated by a comparable language model trained on next-word prediction. Our results suggest that both sentence gestalt update and surprisal predict aspects of N400 amplitudes. Thus, we argue that N400 amplitudes might reflect two distinct but probably closely related sub-processes that contribute to the processing of a sentence.

Nuances of knowing: Brain potentials reveal implicit effects of domain knowledge on word processing in the absence of sentence-level knowledge

In previous work investigating the relationship between domain knowledge (of the fictional world of Harry Potter) and sentence comprehension, domain knowledge had a greater impact on electrical brain potentials to words which completed sentences about fictional "facts" participants reported they did not know compared to facts they did. This suggests that individuals use domain knowledge continuously to activate relevant/related concepts as they process sentences, even with only partial knowledge. As that study relied on subjective reports, it may have resulted in response bias related to an individual's overall domain knowledge. In the present study, we therefore asked participants with varying degrees of domain knowledge to complete sentences describing fictional "facts" as an objective measure of sentence-level knowledge. We then recorded EEG as the same individuals (re-)read the same sentences, including their appropriate final words, and sorted these according to their objective knowledge scores. Replicating and extending Troyer et al., domain knowledge immediately facilitated access to meaning for unknown words; greater domain knowledge was associated with reduced N400 amplitudes for unknown words. These findings constitute novel evidence for graded preactivation of conceptual knowledge (e.g., at the level of semantic features and/or relations) in the absence of lexical prediction. Knowledge also influenced post-N400 memory/integration processes for these same unknown words; greater domain knowledge was associated with enhanced late positive components (LPCs), suggesting that deeper encoding during language processing may be engendered when knowledgeable individuals encounter an apparent gap in their knowledge.

The P600 as a continuous index of integration effort

The integration of word meaning into an unfolding utterance representation is a core operation of incremental language comprehension. There is considerable debate, however, as to which component of the ERP signal-the N400 or the P600-directly reflects integrative processes, with far reaching consequences for the temporal organization and architecture of the comprehension system. Multi-stream models maintaining the N400 as integration crucially rely on the presence of a semantically attractive plausible alternative interpretation to account for the absence of an N400 effect in response to certain semantic anomalies, as reported in previous studies. The single-stream Retrieval-Integration account posits the P600 as an index of integration, further predicting that its amplitude varies continuously with integrative effort. Here, we directly test these competing hypotheses using a context manipulation design in which a semantically attractive alternative is either available or not, and target word plausibility is varied across three levels. An initial self-paced reading study revealed graded reading times for plausibility, suggesting differential integration effort. A subsequent ERP study showed no N400 differences across conditions, and that P600 amplitude is graded for plausibility. These findings are inconsistent with the interpretation of the N400 as an index of integration, as no N400 effect emerged even in the absence of a semantically attractive alternative. By contrast, the link between plausibility, reading times, and P600 amplitude supports the view that the P600 is a continuous index of integration effort. More generally, our results support a single-stream architecture and eschew the need for multi-stream accounts.

An Agent-First Preference in a Patient-First Language During Sentence Comprehension

The language comprehension system preferentially assumes that agents come first during incremental processing. While this might reflect a biologically fixed bias, shared with other domains and other species, the evidence is limited to languages that place agents first, and so the bias could also be learned from usage frequency. Here, we probe the bias with electroencephalography (EEG) in Äiwoo, a language that by default places patients first, but where sentence-initial nouns are still locally ambiguous between patient or agent roles. Comprehenders transiently interpreted nonhuman nouns as patients, eliciting a negativity when disambiguation was toward the less common agent-initial order. By contrast and against frequencies, human nouns were transiently interpreted as agents, eliciting an N400-like negativity when the disambiguation was toward patient-initial order. Consistent with the notion of a fixed property, the agent bias is robust against usage frequency for human referents. However, this bias can be reversed by frequency experience for nonhuman referents.

Fame through surprise: How fame-seeking mass shooters diversify their attacks

Mass shootings are becoming more frequent in the United States, as we routinely learn from the media about attempts that have been prevented or tragedies that destroyed entire communities. To date, there has been limited understanding of the modus operandi of mass shooters, especially those who seek fame through their attacks. Here, we explore whether the attacks of these fame-seeking mass shooters were more surprising than those of others and clarify the link between fame and surprise in mass shootings. We assembled a dataset of 189 mass shootings from 1966 to 2021, integrating data from multiple sources. We categorized the incidents in terms of the targeted population and shooting location. We measured "surprisal" (often known as "Shannon information content") with respect to these features, and we scored fame from Wikipedia traffic data-a commonly used metric of fame. Surprisal was significantly higher for fame-seeking mass shooters than non-fame-seeking ones. We also registered a significant positive correlation between fame and surprisal controlling for the number of casualties and injured victims. Not only do we uncover a link between fame-seeking behavior and surprise in the attacks but also we demonstrate an association between the fame of a mass shooting and its surprise.

Heuristic interpretation as rational inference: A computational model of the N400 and P600 in language processing

Much inquiry in psycholinguistics has focused on evidence from the N400 and P600 components of the event-related potential (ERP) signal-and a central theoretical challenge in this area is accounting for the so-called "semantic P600", which involves unexpected patterns in these components relative to traditional theories of the underlying mechanisms. In this paper we present a computational model of the language processing mechanisms underlying these ERP components, which builds on existing psycholinguistic theories in positing a heuristic interpretation stage of processing, but which deviates from existing theories in formulating this heuristic interpretation process as probabilistic selection via a noisy channel model, and in quantifying and accounting for fine-grained variation in statistical and representational properties of individual stimuli. Our model successfully simulates N400 and P600 patterns from eight psycholinguistic experiments, reflecting the full range of N400-only, P600-only, and biphasic N400-P600 effects, and its behaviors shed light on a number of key patterns that have presented challenges for existing theories. The model's success indicates that a strong account for the processing mechanisms underlying these effects is one in which language comprehension involves a probabilistic heuristic interpretation stage resembling a noisy channel process, feeding into subsequent processes that assess target word fit and reconcile between heuristic and literal interpretations. The model's success also indicates that these mechanisms are critically sensitive to statistical variation in individual stimuli, and that modeling the effects of this variation is essential to account for the full range of observed effects in language processing.

Investigating a neural language model's replicability of psycholinguistic experiments: A case study of NPI licensing

The recent success of deep learning neural language models such as Bidirectional Encoder Representations from Transformers (BERT) has brought innovations to computational language research. The present study explores the possibility of using a language model in investigating human language processes, based on the case study of negative polarity items (NPIs). We first conducted an experiment with BERT to examine whether the model successfully captures the hierarchical structural relationship between an NPI and its licensor and whether it may lead to an error analogous to the grammatical illusion shown in the psycholinguistic experiment (Experiment 1). We also investigated whether the language model can capture the fine-grained semantic properties of NPI licensors and discriminate their subtle differences on the scale of licensing strengths (Experiment 2). The results of the two experiments suggest that overall, the neural language model is highly sensitive to both syntactic and semantic constraints in NPI processing. The model's processing patterns and sensitivities are shown to be very close to humans, suggesting their role as a research tool or object in the study of language.

Is the Mind Inherently Predicting? Exploring Forward and Backward Looking in Language Processing

Prediction is one characteristic of the human mind. But what does it mean to say the mind is a "prediction machine" and inherently forward looking as is frequently claimed? In natural languages, many contexts are not easily predictable in a forward fashion. In English, for example, many frequent verbs do not carry unique meaning on their own but instead, rely on another word or words that follow them to become meaningful. Upon reading take a the processor often cannot easily predict walk as the next word. But the system can "look back" and integrate walk more easily when it follows take a (e.g., as opposed to *make|get|have a walk). In the present paper, we provide further evidence for the importance of both forward and backward-looking in language processing. In two self-paced reading tasks and an eye-tracking reading task, we found evidence that adult English native speakers' sensitivity to word forward and backward conditional probability significantly predicted reading times over and above psycholinguistic predictors of reading latencies. We conclude that both forward and backward-looking (prediction and integration) appear to be important characteristics of language processing. Our results thus suggest that it makes just as much sense to call the mind an "integration machine" which is inherently backward 'looking.'

Native Word Order Processing Is Not Uniform: An ERP Study of Verb-Second Word Order

Studies of native syntactic processing often target phrase structure violations that do not occur in natural production. In contrast, this study examines how variation in basic word order is processed, looking specifically at structures traditionally labelled as violations but that do occur naturally. We examined Swedish verb-second (V2) and verb-third (V3) word order processing in adult native Swedish speakers, manipulating sentence-initial adverbials (temporal idag ‘today’, spatial hemma ‘at home’ and sentential kanske ‘maybe’) in acceptability judgements, in simultaneously recorded event-related potentials (ERP) to visually presented sentences and in a written sentence completion task. An initial corpus study showed that the adverbials differ in frequency in fronted position (idag > kanske > hemma), and although all occur mainly with V2 word order, kanske occurs more frequently with V3 in natural production than both idag and hemma. The experimental results reflected these patterns such that V2 sentences were overall more frequently produced and were deemed more acceptable than V3 sentences. The ERP results consisted of a biphasic N400/P600 response to V3 word order that indicated effects of word retrieval and sentence reanalysis. We also found consistent effects of adverbials. As predicted, V3 was produced more frequently and judged as more acceptable in Kanske sentences than in sentences with the other two adverbials. The ERP analyses showed stronger effects for idag and hemma with V3, especially regarding the P600. The results suggest that the naturally occurring word order ‘violation’, V3 with kanske, is processed differently than V3 with other adverbials where the V2 norm is stronger. Moreover, these patterns are related to individuals’ own production patterns. Overall, the results suggest a more varied native word order processing than previously reported.

Retrieval (N400) and integration (P600) in expectation-based comprehension

Expectation-based theories of language processing, such as Surprisal theory, are supported by evidence of anticipation effects in both behavioural and neurophysiological measures. Online measures of language processing, however, are known to be influenced by factors such as lexical association that are distinct from—but often confounded with—expectancy. An open question therefore is whether a specific locus of expectancy related effects can be established in neural and behavioral processing correlates. We address this question in an event-related potential experiment and a self-paced reading experiment that independently cross expectancy and lexical association in a context manipulation design. We find that event-related potentials reveal that the N400 is sensitive to both expectancy and lexical association, while the P600 is modulated only by expectancy. Reading times, in turn, reveal effects of both association and expectancy in the first spillover region, followed by effects of expectancy alone in the second spillover region. These findings are consistent with the Retrieval-Integration account of language comprehension, according to which lexical retrieval (N400) is facilitated for words that are both expected and associated, whereas integration difficulty (P600) will be greater for unexpected words alone. Further, an exploratory analysis suggests that the P600 is not merely sensitive to expectancy violations, but rather, that there is a continuous relation. Taken together, these results suggest that the P600, like reading times, may reflect a meaning-centric notion of Surprisal in language comprehension.

When components collide: Spatiotemporal overlap of the N400 and P600 in language comprehension

The problem of spatiotemporal overlap between event-related potential (ERP) components is generally acknowledged in language research. However, its implications for the interpretation of experimental results are often overlooked. In a previous experiment on the functional interpretation of the N400 and P600, it was argued that a P600 effect to implausible words was largely obscured - in one of two implausible conditions - by an overlapping N400 effect of semantic association. In the present ERP study, we show that the P600 effect of implausibility is uncovered when the critical condition is tested against a proper baseline condition which elicits a similar N400 amplitude, while it is obscured when tested against a baseline condition producing an N400 effect. Our findings reveal that component overlap can result in the apparent absence or presence of an effect in the surface signal and should therefore be carefully considered when interpreting ERP patterns. Importantly, we show that, by factoring in the effects of spatiotemporal overlap between the N400 and P600 on the surface signal, which we reveal using rERP analysis, apparent inconsistencies in previous findings are easily reconciled, enabling us to draw unambiguous conclusions about the functional interpretation of the N400 and P600 components. Overall, our results provide compelling evidence that the N400 reflects lexical retrieval processes, while the P600 indexes compositional integration of word meaning into the unfolding utterance interpretation.