The Garden Path Sentence Experiments – Thomas Bever The Eye-Tracking During

The human linguistic capacity operates with an astonishing degree of temporal efficiency, transforming serial streams of acoustic disturbances or visual orthographic marks into rich, multidimensional conceptual representations in mere fractions of a second. Under ordinary reading conditions, adult comprehenders identify individual words within 200 to 250 milliseconds, assign grammatical categories, construct intricate hierarchical syntactic trees, and retrieve contextually appropriate semantic interpretations without conscious cognitive effort. Yet, the architectural machinery undergirding this rapid-fire mental parser is not infallible. Precisely because the human language processor operates incrementally—resolving structural ambiguities on a word-by-word basis rather than delaying interpretation until a clause has concluded—it remains susceptible to systematic misdirection. When the parser is lured into committing cognitive resources toward an initially plausible structural analysis that subsequent lexical material proves untenable, the cognitive apparatus experiences a catastrophic processing failure commonly known as a syntactic “garden path.”

The systematic exploration of these structural blind alleys represents one of the foundational triumphs of modern psycholinguistics and cognitive science. Rather than viewing processing breakdowns as peripheral computational errors or idiosyncratic anomalies, researchers recognized that the failure modes of the human parsing mechanism illuminate its latent computational architecture. When a sentence “leads the reader down the garden path,” the observable disruption—manifested as subjective confusion, elevated reading times, regressions to earlier text regions, and localized neural signatures—reveals the invisible, highly constrained heuristics that the mind utilizes to manage real-time informational density. By observing precisely where, how, and why the language processor falters, cognitive scientists can deduce whether the mind constructs syntactic trees serially or in parallel, whether early syntactic decisions are insulated from contextual and pragmatic information, and how working memory interacts with grammatical knowledge.

At the center of this paradigm stands the seminal work of cognitive psychologist Thomas G. Bever, whose 1970 monograph forever altered the trajectory of sentence processing research by formulating the perceptual strategies governing natural language parsing. Bever demonstrated that structural parsing is governed by functional, operational heuristics designed to minimize computational strain, a realization that directly challenged the purely formal, non-temporal models of generative grammar that dominated mid-twentieth-century linguistics. In the decades that followed, the advent of continuous ocular tracking methodologies—spearheaded by researchers such as Keith Rayner, Lyn Frazier, and Marcel Just—transformed Bever’s theoretical hypotheses into empirical, millisecond-by-millisecond behavioral data. By recording the fine-grained oculomotor dynamics of readers as they encounter structural ambiguities, eye-tracking during garden path processing has evolved into an indispensable empirical bedrock for testing competing architectures of the human mind.

1. Introduction to Garden Path Sentences and Thomas Bever’s Foundational Work

1.1 Historical Emergence of Garden Path Phenomena in Psycholinguistics

The mid-twentieth century witnessed a fundamental epistemological shift within the behavioral sciences, marked by the rapid ascendance of the cognitive revolution and the concurrent emergence of psycholinguistics as an autonomous empirical discipline. Central to this transformation was Noam Chomsky’s foundational critique of behaviorist learning models, which replaced stimulus-response contingencies with internal, rule-governed mental representations known collectively as transformational generative grammar. While Chomsky’s formal competence models successfully characterized the abstract mathematical and structural principles governing idealized linguistic systems, they explicitly bracketed the temporal, dynamic, and resource-bounded realities of linguistic performance. Generative grammars operated asynchronously, establishing global relations across hierarchical phrase markers without explicit regard for the biological constraints of immediate human memory, sensory processing limits, or real-time temporal order.

By the late 1960s, a growing cadre of experimentally oriented cognitive psychologists recognized that an adequate theory of language comprehension required an operational model of the human parser—the internal algorithmic processor that translates a linear visual or auditory signal into structural mental representations in real time. Unlike formal competence grammars, which evaluate a completed sentence post hoc, the biological parser is forced to make immediate structural decisions on partial input. Natural language is intrinsically replete with temporary structural ambiguity, wherein a sequence of initial lexical items permits two or more grammatically viable structural continuations. Researchers realized that how the human cognitive apparatus navigates these moments of local syntactic indeterminacy offers an unprecedented window into the fundamental operating system of the human mind.

Psycholinguists quickly observed that parsing failure modes were profoundly more informative than effortless successes. Under normative processing conditions, sentence comprehension is so fluid and automatic that its constituent stages—lexical access, categorical assignment, structural node postulation, and thematic integration—collapse into an apparently seamless, unified percept. However, when the parser encounters a structure that systematically induces a false computational commitment, the seamless illusion fractures. This deliberate manipulation of structural ambiguity revealed that the human parser does not delay analysis until absolute disambiguating evidence emerges. Instead, the mind exhibits a radical bias toward incremental, immediate commitment, even at the perilous cost of being structurally misdirected.

1.2 Thomas Bever’s 1970 Monograph and the Cognitive Revolution

The formal conceptualization of these incremental processing traps crystallized with the publication of Thomas G. Bever’s landmark 1970 monograph, titled The Cognitive Basis for Linguistic Structures. Published within the influential collection edited by John R. Hayes, Bever’s contribution sought to establish a principled boundary—and an interactive bridge—between formal grammatical competence and the operational mechanisms of cognitive psychology. Bever argued that many linguistic universals, constraints on syntactic movement, and distributional properties of human languages do not derive exclusively from innate, encapsulated linguistic faculties, but rather emerge as natural functional adaptations to general human perceptual heuristics and working memory architecture.

To demonstrate the psychological reality of these perceptual heuristics, Bever introduced what has become the most famously parsed, analyzed, and cited specimen in the annals of cognitive psycholinguistics: “The horse raced past the barn fell.” In this canonical construction, the human comprehender universally experiences an acute processing breakdown upon encountering the terminal lexical item “fell.” Bever demonstrated that this failure was not an idiosyncratic consequence of lexical obscurity or grammatical pathology; indeed, the sentence is entirely well-formed according to the standard grammatical rules of English syntax, functioning as a reduced relative clause structurally equivalent to “The horse that was raced past the barn fell.”

Bever leveraged this processing disaster to delineate between formal competence and functional perceptual strategies. He argued that the human mind relies upon rapid, probabilistic heuristic templates to compress incoming perceptual data into meaningful mental structures without incurring crippling cognitive overhead. By illustrating that the human parser treats “The horse raced” as a complete agent-action clause due to ingrained perceptual heuristics, Bever demonstrated that real-time performance strategies impose profound architectural constraints upon linguistic behavior. This foundational work effectively liberated psycholinguistics from the subservient role of merely attempting to validate transformational rules in real time, establishing structural parsing as an independent domain of cognitive psychology dedicated to uncovering the dynamic, heuristic heuristics of the human information processor.

1.3 Defining the Garden Path: Mechanism and Cognitive Disruption

In contemporary psycholinguistic terminology, a garden path sentence is defined as any syntactically well-formed linguistic string that incorporates a localized, temporary structural ambiguity which systematically prompts the human parser to construct an initial syntactic representation that is ultimately incompatible with downstream lexical material. The defining computational characteristic of the garden path phenomenon is serial, incremental commitment. As each lexical token enters the sensory buffer, the parser integrates that token into the emerging hierarchical phrase structure tree immediately, driven by internal simplicity principles, surface-level probabilistic heuristics, or lexical subcategorization biases.

The architecture of a garden path construction consists of three structurally distinct components:

• The Ambiguous Zone: The structural stretch spanning from the onset of the structural ambiguity up to the final element compatible with multiple structural hypotheses (e.g., “The horse raced past the barn…”). Within this zone, the alternative parse (a reduced passive relative clause) remains structurally legitimate, but the parser selectively constructs the dominant alternative (the active main clause).
• The Disambiguating Region: The precise lexical locus—often referred to in the literature as the “crash point” or the disambiguator—that introduces syntactic or morphological material that categorically falsifies the parser’s initial structural hypothesis (e.g., the appearance of the matrix verb “fell”).
• The Reanalysis or Repair Zone: The cognitive and behavioral phase wherein the parser must detect the structural impasse, reject or dismantle the untenable syntactic commitment, search upstream for the structural junction point of misdirection, and reassign lexical elements into their correct hierarchical positions.

The phenomenological experience of encountering a garden path is characterized by an abrupt sensation of cognitive disfluency, conscious surprise, and immediate comprehension failure. Readers frequently report a subjective impression that an entire word or punctuation mark has been omitted from the visual page, or that the sentence is fundamentally ungrammatical. It is essential to emphasize that garden paths are fundamentally distinct from absolute syntactic ill-formedness (e.g., *”The horse fell the barn past raced”*). In an ungrammatical construction, no valid generative parse can be derived from the lexicon and the grammatical rules of the target language. In contrast, a garden path sentence is fully grammatical, possessing a coherent, licensed syntactic structure; its processing difficulty arises entirely from the internal, temporal mechanics of the human processing apparatus, which consistently favors the incorrect structural derivation during real-time, left-to-right processing.

2. Theoretical Foundations: Bever’s Perceptual Strategies and Architectural Hypotheses

2.1 The Canonical Sentence Strategy and NVN Schema

To explain why comprehenders exhibit an overwhelming, deterministic tendency to misanalyze reduced relative constructions, Bever formulated the Canonical Sentence Strategy, historically termed Strategy B. Bever posited that language users internalize dominant surface-structure regularities from natural language input and convert these regularities into rapid, non-modular parsing heuristics. In standard English typography and discourse, the overwhelming statistical majority of simple active declaratives follow a strict canonical sequence: Noun Phrase – Verb – Noun Phrase (NVN). This superficial linear template directly maps onto the fundamental thematic hierarchy of Agent – Action – Patient/Theme.

Bever asserted that when the human parser encounters an initial sequence consisting of a potential nominal followed immediately by a verb capable of active transitive or intransitive interpretation, it automatically applies the NVN perceptual template:

[NP1 … V … (NP2)] → [Agent … Action … Patient]

This mapping operates as a cognitive economizer, allowing the parser to assign structural roles and thematic values without waiting to evaluate whether the verb might function as a passive past participle introducing a reduced relative clause. When reading “The horse raced…”, the parser classifies “The horse” as the grammatical subject and thematic Agent, and “raced” as the active finite matrix predicate. The prepositional phrase “past the barn” is effortlessly integrated as an adverbial locative adjunct modifying the matrix event of racing.

Extensive empirical investigations across developmental psycholinguistics have confirmed the profound cognitive entrenchment of this strategy. Young children, whose cognitive control mechanisms and working memory capacities are still maturing, demonstrate an extreme, unshakeable over-reliance on the NVN heuristic. When presented with complex non-canonical sentences such as passive constructions (“The horse was kicked by the cow”), young children frequently misinterpret the first noun as the Agent, reporting that the horse kicked the cow. In adults, while syntactic competence is fully developed, the primary, millisecond-by-millisecond parsing heuristic continues to default to this surface-level linear template, establishing the initial misparse that makes subsequent structural reanalysis computationally mandatory.

2.2 The Principle of Late Closure and Minimal Attachment Antecedents

Although Bever framed his theoretical accounts in terms of surface-level perceptual strategies and communicative heuristics, his observational analyses laid the structural groundwork for the formal algorithmic parsing models developed in subsequent decades. Most notably, Bever observed that human comprehenders consistently attempt to integrate incoming lexical material into the structural unit or clausal boundary that is currently undergoing active compilation in short-term memory, rather than postulating a higher, overarching clausal node that would require closing the active constituent.

This perceptual observation directly anticipated Lyn Frazier’s subsequent formalization of the Principle of Late Closure within the Garden-Path Model. Frazier defined this principle as an architectural imperative: if grammatically permissible, attach new incoming lexical items into the clause or phrase currently being processed. By maintaining structural integration within the local phrase marker, the cognitive parser drastically reduces processing load by preventing the premature closure of working memory buffers. In parallel, Bever’s conceptualization of perceptual simplicity anticipated Frazier’s complementary metric, the Principle of Minimal Attachment, which mandates that the parser construct the structurally simplest analysis licensed by the grammar, postulating the absolute minimum number of internal syntactic nodes.

The continuity from Bever’s operational perceptual strategies to Frazier’s structural parsing heuristics marks a crucial theoretical evolution. While Bever viewed these mechanisms as pragmatic, heuristic short-cuts designed to cope with real-time perceptual flux, Frazier and her collaborators abstracted these tendencies into universal, domain-specific geometric principles of structural tree construction. Both frameworks, however, share the fundamental architectural premise that the initial structural commitments of the human mind are governed by strict cognitive economization constraints that prioritize computational velocity and local economy over exhaustive structural search.

2.3 Modularity vs. Interactionism in Early Parsing Theory

Bever’s monograph served as an intellectual catalyst for one of the most vigorously contested debates in the cognitive sciences: the architectural conflict between Modularity and Interactionism. The core theoretical question asks whether the human sentence comprehension mechanism operates via an encapsulated, autonomous syntactic subcomponent that performs structural assignments in complete isolation from contextual, pragmatic, and semantic real-world knowledge, or whether all available sources of linguistic and extra-linguistic information interact continuously and simultaneously to constrain syntactic interpretation from the earliest millisecond of processing.

The modular framework—philosophically articulated by Jerry Fodor and operationalized in parsing theory by Lyn Frazier and Keith Rayner—posited a strictly serial, two-stage architecture. Under this model, an encapsulated “syntactic reflex” operates first, utilizing only lexical category information and structural simplicity metrics (Minimal Attachment and Late Closure) to construct a single syntactic candidate tree. Only after this initial, non-interactive syntactic pass does a secondary semantic and pragmatic filter evaluate the output; if the structural analysis clashes with contextual semantics or world knowledge, a computationally costly reanalysis routine is triggered. Bever’s demonstrations of garden path phenomena were frequently marshaled as prime empirical evidence for modularity: if higher-order world knowledge (e.g., knowing that barns do not walk or that horses are frequently driven by human agents) were immediately integrated, the parser should theoretically resist the garden path in contexts where semantic plausibility disfavors the active main-verb interpretation.

Conversely, early interactionist theorists argued that Bever’s perceptual strategies were not encapsulated syntactic algorithms, but rather probabilistic manifestations of broader cognitive heuristics that integrate semantic plausibility, discourse context, and communicative intent simultaneously. The central computational challenge for early interactionist models was accounting for the indisputable temporal persistence of garden path errors: if the parser possesses simultaneous access to semantic and contextual data, why does it consistently fall victim to structural misdirection even when the target structure is syntactically coherent and contextually predictable? This dialectic established the empirical agenda for the following four decades of psycholinguistic experimentation, driving the development of increasingly refined temporal methodologies capable of dissecting the microsecond-level mechanics of syntactic comprehension.

3. Structural Mechanics of Classic Garden Path Constructions

3.1 Reduced Relative Clauses vs. Main Verb Ambiguities

The morphological and syntactic architecture of the English language creates fertile structural terrain for garden path phenomena, primarily due to the widespread phenomenon of morphological syncretism between the simple past tense (preterite) and the past participle forms of regular verbs (and a substantial class of irregular verbs). In standard orthography, the addition of the -ed suffix marks both the active finite preterite (e.g., “The horse raced” [intransitive or active transitive]) and the passive non-finite past participle (e.g., “The horse [which was] raced” [passive participle]).

When the parser encounters a string configured as:

[Determiner + Noun + Verb-ed + Prepositional Phrase + Finite Verb]

two fundamentally distinct syntactic derivations compete for assignment within the emerging phrase structure tree:

1. The Main Clause Hypothesis: The initial Noun Phrase is projected as the external argument (Subject/Agent) of the matrix clause, the Verb-ed form is assigned the status of matrix finite verb, and the following Prepositional Phrase is parsed as an adjunct dependent upon that matrix verb.
2. The Reduced Relative Clause Hypothesis: The initial Noun Phrase is projected as the matrix subject, but the Verb-ed form is analyzed as the head of a non-finite participial modifier clause (an adjectival adjunct) containing an elided passive relativizer and auxiliary (e.g., “which was”), with the nominal functioning as the internal argument (Theme/Patient) of the participial verb.

The insertion of explicit overt grammatical morphology—such as the relative pronoun and auxiliary in “The horse that was raced past the barn fell”—completely abolishes the ambiguity, forcing the immediate projection of an embedded clausal node. However, when these elements are phonologically and orthographically suppressed via standard English reduction rules, the parser’s structural minimization bias categorically selects the Main Clause Hypothesis. The phrase structure tree generated during this initial parse fails to provide an empty structural slot to license the downstream appearance of the genuine matrix finite verb (“fell”). Consequently, when “fell” is encountered, it functions as an unlicensed, syntactically illicit second predicate within a clause that has already saturated its verb valency, generating an absolute syntactic impasse that demands total structural re-evaluation.

3.2 Direct Object vs. Complement Clause Ambiguities

A second canonical class of garden path constructions involves structural competition between a simple Direct Object noun phrase and a Sentential Complement clause, frequently occurring after verbs of cognition, perception, or communication (e.g., know, believe, realize, suspect, admit). Consider the classic paradigm:

“The student knew the answer was incorrect.”

In this syntactic configuration, the lexical verb “knew” possesses a flexible subcategorization frame; it can either license a simple nominal direct object (NP complement: “The student knew [NP the answer]”) or an entire embedded declarative proposition (Sentential Complement: “The student knew [CP (that) [S the answer was incorrect]]”).

When the parser encounters the post-verbal nominal constituent “the answer”, it faces an immediate attachment decision before reading the rest of the sentence. Guided by the Principle of Minimal Attachment, the parser projects a simple Direct Object relationship, directly attaching the NP under the primary Verb Phrase (VP) node:

[S [NP The student] [VP [V knew] [NP the answer]]]

This syntactic structure requires fewer phrase-structural nodes than constructing an abstract Complementizer Phrase (CP) containing a subordinate sentence (S/IP) node with an empty complementizer head. Furthermore, the parser assigns an immediate thematic role to “the answer”, designating it as the direct Theme or cognitive object of the student’s knowledge state.

The cognitive disruption occurs precisely when the parser encounters the subsequent copular verb “was”. At this junction, the direct object analysis is categorically invalidated: a simple nominal direct object cannot be followed immediately by a finite copula and predicate adjective without an intervening syntactic coordinator. The parser is forced to retroactively alter the structural category of “the answer” from an internal direct object of the matrix verb into the structural subject of an embedded complement clause. When the overt complementizer “that” is present (“The student knew that the answer was incorrect”), the ambiguity is structurally pre-empted; its systematic omission in natural speech and typography guarantees that comprehenders frequently experience localized garden-pathing at the subordinate auxiliary or predicate.

3.3 Prepositional Phrase Attachment and Modifier Ambiguities

Prepositional Phrase (PP) attachment ambiguities represent a structurally intricate category of garden path sentences that uniquely highlights the spatial geometric constraints governing the mental parser. A classic illustrative specimen is:

“The cop saw the spy with the binoculars.”

Here, the incoming prepositional phrase “with the binoculars” permits two legitimate structural attachments within the hierarchical syntactic tree:

1. High Attachment (VP-Attachment): The PP attaches directly to the higher Verb Phrase node dominated by “saw”. In this configuration, the prepositional phrase functions as an instrumental modifier characterizing the perceptual action of the subject: the police officer utilized binoculars to execute the visual observation.
2. Low Attachment (NP-Attachment): The PP attaches locally to the internal Noun Phrase node dominated by “the spy”. In this configuration, the prepositional phrase acts as a restrictive or non-restrictive nominal modifier: the spy was in possession of binoculars, and the officer observed this particular spy.

According to the Principle of Minimal Attachment, the parser exhibits an automatic default bias toward High Attachment (VP-attachment) because attaching the PP directly to the preexisting VP node requires the postulation of fewer additional hierarchical structural nodes than expanding the nominal complex with an internal NP-modifier structure. However, the structural dynamics become significantly more volatile when lexical and semantic biases conflict with this structural default. In sentences such as:

“The cop saw the spy with the revolver.”

the instrumental reading (High Attachment) implies that the cop utilized a firearm to execute a visual observation—a pragmatically bizarre interpretation that clashes with real-world semantic frames. The parser is thus forced to mediate between structural simplicity metrics that favor High Attachment and lexical-pragmatic plausibility constraints that mandate Low Attachment, frequently resulting in observable processing delays, fixation regressions, and cognitive reanalysis overhead.

4. The Emergence of Eye-Tracking Methodology in Psycholinguistics

4.1 From Self-Paced Reading to Continuous Eye-Tracking

Before the widespread empirical deployment of high-precision ocular tracking instrumentation, the primary behavioral method for studying sentence comprehension in real time was the Self-Paced Reading (SPR) paradigm. In a moving-window self-paced reading experiment, a participant sits before a monitor displaying a series of masked hyphens or underscores hiding words. By pressing a physical button, the participant unmasks a specific word or region of text while simultaneously masking the preceding item. The latency between successive button presses is recorded as an operational index of processing difficulty at that specific structural locus.

While self-paced reading provided valuable early insights, it suffered from profound methodological and ecological limitations:

• Motor Artifacts: The measured latency reflects not only cognitive parsing and lexical access, but also motor planning, neural efference to the fingers, and mechanical actuation of the physical switch, introducing significant noise into the temporal data.
• Unnatural Pacing: The paradigm destroys the continuous, holistic visual flow of natural reading, artificially impeding normal reading dynamics and preventing readers from developing normal parafoveal previews.
• Impaired Regressive Capability: Standard self-paced designs preclude or severely penalize natural regressive saccades. Readers cannot smoothly glance backward to inspect earlier words; they are either entirely forbidden from regressing (in non-cumulative designs) or must perform awkward, serial button presses to uncover upstream text, radically distorting the reanalysis process.

The widespread adoption of continuous, non-invasive eye-tracking systems in the late 1970s and early 1980s revolutionized reading research. Led by pioneering experimentalists such as Keith Rayner, eye-tracking paradigms captured the natural, unrestrained ocular movements of human comprehenders reading running text at native speeds. By measuring the reflection of infrared light off the cornea and the center of the pupil, modern eye-trackers record ocular fixations and saccades with sub-millisecond temporal resolution and sub-character spatial accuracy, providing an ecological, continuous window into the immediate operations of the human linguistic processor.

4.2 The Eye-Mind Hypothesis and Cognitive Transparency

The epistemological foundation that licenses the theoretical inference from ocular behavior to internal cognitive processes is known as the Eye-Mind Hypothesis (or the Eye-Mind Assumption), systematically articulated by Marcel A. Just and Patricia A. Carpenter in their landmark 1980 psychological review. The Eye-Mind Hypothesis asserts that there is no appreciable lag between the ocular fixation on a particular visual item and the cognitive processing of that item: the eye remains fixated on a word for as long as that word is being actively processed by the cognitive apparatus, and shifts to a new word only when the cognitive processing of the currently fixated item has completed or reached an operational threshold.

Formally, the Eye-Mind framework assumes two interconnected operational axioms:

1. Immediacy of Processing: The human parser attempts to interpret each incoming word immediately at all levels of the linguistic hierarchy—executing orthographic decoding, lexical access, syntactic categorical assignment, semantic role integration, and referential assignment—without waiting for structural or lexical resolution from downstream context.
2. Temporal Co-registration: The duration of an ocular fixation directly reflects the internal computational complexity encountered by the mental processor while handling the visual and conceptual properties of that specific orthographic token.

Although decades of subsequent research have revealed critical caveats to this assumption—most notably the existence of “spillover effects” (wherein computational difficulty initiated at word N delays processing during fixation on word N+1) and the pre-processing of parafoveal information—the Eye-Mind Hypothesis remains a foundational methodological cornerstone. It provides the explicit empirical justification for utilizing ocular metrics as a direct, unmediated readout of instantaneous syntactic parsing difficulty during the processing of garden path sentences.

4.3 Apparatus Evolution: From Scleral Coils to Modern Infrared Optics

The technological trajectory of ocular measurement in experimental psychology represents a transition from invasive mechanical intervention to pristine optical engineering. In the earliest twentieth-century studies of reading, researchers utilized invasive contact lenses known as scleral search coils. These devices required participants to insert a physical lens directly onto the eye, held in place by vacuum suction, with tiny physical levers or electromagnetic induction coils running from the eye to mechanical or magnetic recording apparatuses. While providing exceptional spatial accuracy, these systems caused substantial physical discomfort, restricted experimental session durations, introduced severe ocular fatigue, and radically distorted the normative reading experience.

A transformative breakthrough occurred with the development of the Dual-Purkinje Image (DPI) optical tracking system. The DPI tracker operates by projecting a beam of infrared light toward the ocular surface and continuously measuring the spatial relationship between two distinct optical reflections: the First Purkinje Image (the bright reflection off the anterior surface of the cornea) and the Fourth Purkinje Image (the reflection off the posterior surface of the crystalline lens inside the eye). Because the corneal and lens reflections shift differentially during ocular rotation but move identically during translational head movements, DPI systems permitted high spatial resolution without requiring physical attachments to the globe of the eye.

In contemporary psycholinguistic laboratories, experimental research is dominated by video-based, infrared corneal-reflection eye-trackers, manufactured by specialized engineering firms such as SR Research (the EyeLink systems). These devices utilize high-speed digital video cameras capturing ocular images at sampling frequencies ranging from 1,000 to 2,000 Hertz (one to two observations per millisecond). A stabilized infrared illuminator creates a distinct corneal reflection (glint) while illuminating the pupil. Advanced machine-vision algorithms isolate the center of the dark pupil and the glint vector in real time, compensating automatically for subtle head drifts. These systems maintain spatial tracking accuracy within 0.25 to 0.5 degrees of visual angle—sufficient to reliably isolate whether a reader is fixating on a specific letter within a critical disambiguating word.

5. Key Eye-Tracking Metrics in Sentence Processing

5.1 Early Stage Processing Metrics: First Fixation and First-Pass Time

When analyzing eye-tracking data collected during sentence comprehension experiments, researchers divide ocular behavior across specific spatial Regions of Interest (ROIs) into early-stage and late-stage temporal metrics. This chronological dissociation is of paramount theoretical significance: early-stage metrics capture the initial, feed-forward sweep of visual intake, lexical identification, and initial structural node attachment, whereas late-stage metrics capture downstream integrative processes, error detection, and structural reanalysis routines.

The primary early-stage eye-tracking metrics comprise:

• First Fixation Duration (FFD): The temporal duration (in milliseconds) of the absolute first ocular fixation landed within a targeted Region of Interest, provided that the reader has arrived at this region from the left without prior exposure. FFD is widely considered an uncontaminated index of low-level visual processing and initial lexical access. In garden path experiments, if an experimental manipulation affects FFD at the disambiguating word, it indicates that the structural incompatibility is recognized with extraordinary, reflexive immediacy.
• Gaze Duration (also termed First-Pass Reading Time): The sum of all fixation durations within a targeted Region of Interest from the moment the eye first enters the region until the gaze leaves that region in any direction (either progressing forward to downstream text or regressing backward to upstream text). Gaze duration is the gold-standard metric for assessing initial syntactic parsing commitments. If the mental parser operates via modular, syntax-first heuristics, a statistically significant inflation in First-Pass Reading Time will emerge on the critical disambiguating verb, reflecting the localized cognitive shock triggered by the structural mismatch.
• Single Fixation Duration (SFD): The duration of a fixation within a region when that region receives precisely one single fixation before the gaze moves onward. SFD provides a pure metric of uncomplicated lexical integration, though it necessarily excludes trials where complex reanalysis immediately precipitates multiple fixations.

5.2 Late Stage Processing Metrics: Regressions and Total Time

While early-stage metrics illuminate initial structural commitments, late-stage eye-tracking metrics capture the overt behavioral consequences of parsing failure—specifically, the cognitive struggle to diagnose, dismantle, and repair an untenable syntactic representation. When the human parser crashes upon encountering an illicit lexical item within a garden path structure, the ocular motor program halts forward progress and executes corrective maneuvers.

The classic late-stage processing metrics include:

• Regressions Out (Regression Probability): The proportion of trials in which the reader’s gaze executes an initial regressive saccade (a backward jump to the left) leaving the target region, before any forward saccade is initiated. In garden path sentences, the disambiguating region typically exhibits a dramatic, explosive spike in Regressions Out (often exceeding 50% to 70% of trials), providing unequivocal behavioral confirmation of cognitive processing failure.
• Regression Path Duration (also termed Go-Past Time): The total elapsed time from the instant the eye first enters the target region until the eye moves past the target region to the right, into new downstream material. Critically, this metric includes all the time spent executing regressive saccades back into upstream regions, re-reading preceding text, and returning to the target word. Go-Past time is arguably the most sensitive oculomotor measure of reanalysis effort: it quantifies the precise temporal tax incurred by the reader to resolve the garden path and successfully license the disambiguating item.
• Second-Pass Reading Time: The sum of all fixation durations spent within a previously visited region after the gaze has regressed back into that region from downstream text. Elevated second-pass times over the subject noun phrase or the ambiguous verb indicate that the reader is actively inspecting those earlier anchor points to construct an alternative syntactic tree.
• Total Reading Time: The absolute summation of every fixation duration committed to a specific region across the entire reading trial. Total time provides an overarching, holistic index of cumulative processing difficulty, integrating early lexical recognition, initial parsing costs, regressive re-reading passes, and global semantic integration.

5.3 Spatial Saccadic Dynamics and Skipping Rates

Beyond temporal duration measurements, the spatial characteristics of ocular movements provide essential data regarding the predictability and difficulty of linguistic structures. During native reading, readers do not fixate every single word sequentially; approximately 30% of words in continuous text are entirely skipped, with the eye jumping over them directly to subsequent lexical tokens. Word Skipping Rate is heavily modulated by lexical length and contextual predictability: short function words (e.g., “the,” “of,” “in”) and highly predictable content words are routinely skipped, whereas low-frequency or structurally unexpected words consistently demand direct ocular fixation.

In garden path paradigms, spatial dynamics manifest in several distinct ways:

• Parafoveal Processing and the Boundary Paradigm: Human visual acuity is partitioned into foveal vision (the central 1–2 degrees of the visual field where acuity is sharpest), parafoveal vision (extending symmetrically roughly 5 degrees around the fovea), and peripheral vision. Utilizing Keith Rayner’s ingenious Moving-Window and Gaze-Contingent Boundary Paradigms, researchers demonstrated that readers extract substantial orthographic, phonological, and morphological information from the upcoming word (parafoveal preview) before the eye ever lands upon it. When approaching a disambiguating word in a garden path sentence, an incongruous parafoveal preview can trigger pre-fixation saccadic slowing or inflate the duration of the fixation immediately preceding the target region.
• Saccade Length and Target Landing Sites: Forward saccades in typical English reading average roughly 7 to 9 character spaces in length. However, when the parser encounters a syntactic crash at a disambiguating word, forward saccade lengths compress dramatically or cease altogether. More importantly, the spatial trajectories of regressive saccades provide direct empirical clues regarding the nature of the cognitive repair mechanism: regressions do not jump blindly into visual space, but frequently land with remarkable spatial precision upon the ambiguous verb or the initial nominal phrase, indicating that the mind maintains an accurate spatial and syntactic map of the antecedent text during parsing collapse.

6. Frazier and Rayner’s Experimental Validations of Bever’s Hypotheses

6.1 The Two-Stage Garden-Path Model Formalization

Building upon the foundations laid by Thomas Bever, psycholinguists Lyn Frazier and Keith Rayner conducted a series of landmark eye-tracking experiments that culminated in their seminal 1982 publication in Cognitive Psychology, titled “Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences.” Frazier and Rayner sought to transform Bever’s intuitive perceptual strategies into a rigorous, predictive, and mathematically explicit algorithmic theory: the Garden-Path Model (often designated as the Frazier-Rayner Two-Stage Architecture).

Frazier and Rayner formalized two structural heuristics that operate unconditionally during the initial parsing pass:

1. The Minimal Attachment Principle: The parser attachments incoming lexical items into the emerging phrase structure tree using the fewest possible structural nodes, strictly adhering to the simplest hierarchical geometric representation licensed by the grammar.
2. The Late Closure Principle: If Minimal Attachment provides no preference (i.e., alternative structural analyses postulate an identical number of additional syntactic nodes), incoming lexical items must be attached directly into the clause or phrase currently undergoing active processing in working memory.

Frazier and Rayner framed these principles as an explicit evolutionary and computational necessity. Because the human brain possesses finite working memory buffers and a limited sensory store, it cannot afford to maintain dozens of competing parallel syntactic derivations simultaneously without incurring massive computational slowdowns. By executing an immediate, deterministic, serial structural commitment governed by simplicity principles, the parser minimizes cognitive load and achieves instantaneous interpretation. Bever’s classic “The horse raced past the barn fell” was thus revealed not as an idiosyncratic cognitive error, but as the inevitable, mathematically predictable output of an elegant, serial computational engine applying Minimal Attachment to avoid processing delays.

6.2 Oculomotor Manifestations at the Disambiguating Region

To provide empirical proof for their Two-Stage model, Frazier and Rayner tracked the real-time ocular fixations of participants reading syntactically matched pairs of sentences containing temporary structural ambiguities, contrasted with unambiguous control conditions. For instance, they examined direct object versus sentential complement ambiguities (e.g., “Since Jay always jogs a mile seems like a short distance to him” versus “Since Jay always jogs, a mile seems like a short distance to him”) and reduced relative versus full relative constructions.

The eye-tracking results provided stunning, incontrovertible vindication of Bever’s predictions and Frazier’s algorithmic formalization:

• Immediate Disambiguation Disruption: In the garden path conditions, readers exhibited normal, fluent reading speeds through the entire ambiguous zone. However, the instant the ocular gaze landed upon the disambiguating region (e.g., the word “seems” or “fell”), the eye-movement record exhibited an immediate, catastrophic disruption. First Fixation Durations and Gaze Durations on the disambiguating word spiked significantly compared to unambiguous controls.
• Absence of Delay: The parser did not adopt a “wait-and-see” strategy. If the parser had delayed its structural decision, reading times in the ambiguous region would have been elevated compared to controls, and no disruption would have emerged at the disambiguator. The immediate, localized disruption on the disambiguating word proved that the human parser constructs syntactic commitments incrementally, word by word, without the slightest temporal hesitation.
• Explosion of Regressive Saccades: The probability of executing a backward saccade (Regressions Out) escalated exponentially at the disambiguating token. Readers who read smoothly through the ambiguous region suddenly halted forward movement and launched regressions back toward the ambiguous verb and the initial subject noun phrase, providing an observable, objective trace of the internal cognitive crash.

6.3 Evidence of Reanalysis Cost and Trajectory

Frazier and Rayner’s 1982 experiments went far beyond merely demonstrating that garden path sentences are difficult; they provided the first empirical quantification of the cost and trajectory of syntactic reanalysis. By tracking the precise spatial coordinates where regressive saccades landed after departing the disambiguating region, the authors were able to reconstruct the mental steps taken by readers attempting to rescue a failed parse.

The ocular trajectories revealed two distinct phases of recovery:

1. Targeted Backtracking: When the parser encountered the disambiguating verb, the initial regressive saccade did not drift randomly across the monitor. Instead, regressions exhibited a powerful spatial concentration targeting the precise location of the ambiguous verb (e.g., “raced”) or the clause boundary. This targeted oculomotor behavior demonstrated that the cognitive architecture retains an indexed memory of structurally critical pivot points, allowing the reanalysis routine to jump directly to the structural node requiring revision.
2. The Reanalysis Temporal Tax (Go-Past Inflation): Frazier and Rayner observed massive inflations in Regression Path Duration (Go-Past Time) for garden path sentences. Readers spent hundreds of additional milliseconds oscillating their ocular gaze between the ambiguous verb, the modifying prepositional phrase, and the disambiguating predicate. Only after this cyclic, second-pass re-inspection did the gaze finally move past the disambiguating word into the remainder of the sentence. This empirical duration directly quantified the internal energetic cost of dismantling the Minimal Attachment tree and projecting the more complex, non-minimal reduced relative structure.

7. Syntactic Disambiguation and the Mechanics of Cognitive Reanalysis

7.1 Diagnosis of Structural Incompatibility

What specific computational mechanisms trigger the parsing crash when the eye lands on a disambiguating lexical item? In the theoretical architecture of the human parser, syntactic disambiguation represents a moment of acute feature mismatch. The human linguistic processor continuously maintains an active representation of expectations: based on the currently constructed phrase marker, the parser anticipates specific grammatical categories, morphological features, and thematic arguments to satisfy open structural licenses.

When reading a sentence like “The defendant examined by the lawyer was guilty,” the parser implements Bever’s NVN strategy, projecting “The defendant examined…” as an active transitive clause where “examined” is the matrix verb requiring a direct object (Theme). Upon encountering the preposition “by”, the parser receives an immediate structural shock:

• Structural Stall: The preposition “by” cannot satisfy the active direct object requirement of the verb “examined”. Instead, “by” functions as an agentive prepositional marker characteristic of passive constructions.
• Categorical Invalidation: In the classic Bever sentence, “The horse raced past the barn fell,” the crash is even more definitive. The parser arrives at “fell” having already constructed a complete, well-formed, and structurally saturated main clause: Subject = “The horse”; Finite Verb = “raced”; Locative Adjunct = “past the barn”. The incoming lexical item “fell” is unambiguously classified as an intransitive finite verb. The parser possesses no open structural slot, no unfilled syntactic position, and no unsatisfied subcategorization frame capable of licensing this second finite verb.

At this microsecond interval, the internal parser experiences an unresolvable categorical block. The computational system cannot link the incoming lexical head to the existing root node without violating fundamental syntactic principles (such as the requirement that a single clause have only one finite matrix predicate unless connected by coordination or subordination). This diagnostic recognition triggers an immediate interrupt signal within the cognitive architecture, terminating normal feed-forward reading and initiating an emergency syntactic repair routine.

7.2 Models of Syntactic Repair: Selective Reanalysis vs. Complete Restart

Once structural incompatibility has been diagnosed, the parser must execute a repair. Psycholinguists have proposed fundamentally contrasting theoretical models regarding how the cognitive apparatus accomplishes this structural restoration, which have been vigorously evaluated using eye-movement trajectory data:

The primary theoretical models of syntactic repair include:

• The Complete Restart Model (Backtracking to Root): This early hypothesis proposed that upon encountering a syntactic crash, the parser wipes its short-term working memory clean, abandons the entire constructed phrase marker, returns its ocular gaze to the absolute beginning of the sentence (the initial word), and initiates an entirely new, exhaustive, left-to-right parsing attempt, this time suppressing the initially favored structural alternative.
• The Selective Reanalysis Hypothesis: Proposed by Frazier and Rayner (1982), this model asserts that the parser is a highly efficient, targeted repair engine. Rather than destroying the entire syntactic tree, the reanalysis mechanism uses the specific categorical identity of the disambiguating item to diagnose the exact node where the initial misparse occurred. The gaze executes a targeted saccadic jump directly to that structural nexus—bypassing intervening words—and performs a localized structural adjustment (e.g., retroactively inserting a non-finite relative clause node above the ambiguous verb).
• The “Attach Anyway” and “Adjust” Diagnostic Repair Model: Formulated by Janet Dean Fodor and Atsu Inoue (1994, 1998), this influential computational model posits that the human parser never completely restarts and rarely engages in abstract, non-local search. Instead, when the parser encounters an incompatible word, it operates under the principle of Attach Anyway: it forces the incoming word into the existing phrase marker at the most plausible local site, even if that attachment violates grammatical constraints. This localized attachment creates an acute structural strain or error. The parser then executes the Adjust phase: it uses the localized grammatical violation as an informational symptom to mechanically shift or “repair” adjacent structural nodes until grammatical equilibrium is restored.

The eye-tracking evidence overwhelmingly refutes the Complete Restart Model. Readers rarely regress all the way to the first word of a sentence unless the garden path is extraordinarily severe or their working memory has been completely overwhelmed. Instead, the oculomotor trace reveals highly selective, targeted landing sites centered precisely on the ambiguous verb and the clausal junction, providing definitive empirical validation for targeted, diagnostic repair models.

7.3 Lexical and Semantic Coercion During Syntactic Correction

Syntactic reanalysis is not an abstract, purely geometric exercise; it imposes severe cognitive processing demands because it requires the human brain to execute rapid lexical and semantic coercion. When the parser resolves a garden path like “The horse raced past the barn fell,” the internal mental model of the event must undergo radical restructuring:

This cognitive overhaul involves several concurrent computational steps:

1. Morphological and Categorical Coercion: The lexical entry for “raced” must be forcibly re-evaluated. Its categorical assignment must be stripped of its finite, active past-tense status and coerced into a non-finite, passive past participle.
2. Thematic Role Reassignment: In the initial parse, the mental model constructed an Agent role for “The horse” (the horse is actively executing the physical motion of galloping). During reanalysis, this thematic assignment must be erased and overhauled: “The horse” must be demoted to the Theme/Patient role (the horse is an entity being ridden, driven, or raced by an implicit, unstated human agent).
3. Suppression of the Active Schema: The pre-activated semantic frame of an independent, self-propelled event must be actively suppressed in working memory, while a complex, embedded, multiclause proposition is retrieved and instantiated.

This multi-level semantic reassignment explains why garden path repair generates such immense cognitive friction. Individual differences in Working Memory Capacity (typically measured via reading span tasks) strongly modulate the velocity and success of this reanalysis process. Readers with high working memory capacity possess the cognitive resources necessary to simultaneously maintain the old structural representation while compiling the new coerced representation, executing targeted regressions and recovering within milliseconds. Readers with lower working memory capacity frequently experience total reanalysis failure, getting trapped in cyclic, chaotic ocular regressions across the entire visual display or failing to extract the correct semantic meaning altogether.

8. The Garden-Path Model vs. Constraint-Satisfaction Approaches

8.1 The Constraint-Based Interactive Alternative

Throughout the late 1980s and 1990s, the modular, two-stage Garden-Path Model was radically challenged by the emergence of Constraint-Based Interactive Models of sentence processing, formulated by cognitive scientists such as Maryellen C. MacDonald, Mark S. Seidenberg, John C. Trueswell, and Michael K. Tanenhaus. Drawing inspiration from connectionist, parallel-distributed processing (PDP) architectures and statistical linguistics, constraint-based theorists rejected the concept of an encapsulated, syntax-first parsing reflex.

Instead, the constraint-satisfaction framework posits that:

• Parallel Continuous Activation: The human parser maintains multiple competing structural analyses simultaneously within a continuous, dynamic activation landscape.
• Simultaneous Multi-Source Constraints: Syntactic alternative derivations compete for activation, and this competition is modulated from the earliest millisecond of processing by every available source of information, including lexical frequency, verb argument preferences, semantic plausibility, discourse context, referential visual cues, and prosodic boundaries.
• Absence of Catastrophic Structural Crash: Under this view, what appears to be a sudden “garden path crash” is simply the cognitive friction of resolving acute competition between two highly activated, mutually inhibitory structural alternatives. If contextual, lexical, and semantic constraints heavily favor the subordinate or reduced relative interpretation from the outset, the parser should process the sentence without any garden path effect whatsoever.

8.2 Lexical Bias and Verb Transitivity Entailments

The primary empirical battleground between the Garden-Path Model and Constraint-Based models centered on the role of lexical argument structure biases. Proponents of constraint satisfaction pointed out that verbs differ profoundly in their statistical usage in natural language corpora: some verbs are overwhelmingly transitive (e.g., “hit,” “punish”), others are strongly intransitive (e.g., “sleep,” “arrive”), and ambiguous verbs exhibit distinct probabilistic preferences for either active past-tense forms or passive participial forms (e.g., “selected” vs. “raced”).

In a series of landmark eye-tracking experiments, John Trueswell, Michael Tanenhaus, and Susan Garnsey systematically tested whether lexical bias could eliminate garden path effects. Garnsey and colleagues (1997) presented readers with Sentential Complement ambiguities using matrix verbs that had either a strong Direct Object bias (e.g., “The senior advisor confirmed the rumor was completely false”) or a strong Sentential Complement bias (e.g., “The senior advisor claimed the rumor was completely false”).

The eye-tracking metrics revealed striking divergence:

• When the verb possessed a strong Direct Object bias, readers fell deeply into the garden path at the disambiguating auxiliary “was”, exhibiting classic inflations in First-Pass Reading Time, massive Regression Path Durations, and high regression rates.
• However, when the verb possessed a strong Sentential Complement bias, the garden path disruption at “was” was significantly attenuated or entirely eliminated: first-pass fixations on the auxiliary showed no statistically significant elevation compared to unambiguous control sentences containing the overt complementizer “that”.

These findings provided formidable empirical support for constraint-based architectures. They demonstrated that the mental parser is exquisitely tuned to the statistical, fine-grained lexical properties of individual verbs. The human parser does not rigidly follow blind geometric rules like Minimal Attachment; rather, it rapidly deploys probabilistic lexical knowledge to guide initial structural activation, preventing garden paths whenever local lexical constraints provide clear predictive support for the structurally more complex parse.

8.3 Visual World Paradigm and Referential Context

The definitive empirical challenge to modular parsing theory arrived with the invention of the Visual World Paradigm by Michael K. Tanenhaus, Michael J. Spivey-Knowlton, Kathleen M. Eberhard, and Julie C. Sedivy in 1995. Published in Science, their landmark study titled “Integration of visual and linguistic information in spoken language comprehension” adapted high-speed eye-tracking to record the ocular gaze of participants looking at physical objects in real-world visual displays while listening to spoken auditory instructions.

Tanenhaus and colleagues investigated temporary prepositional phrase attachment ambiguities using instructions such as:

“Put the apple on the towel in the box.”

In an isolated, text-based environment, Minimal Attachment dictates that “on the towel” must be parsed as the Goal argument of “put” (High VP-attachment), inducing a catastrophic garden path when the listener encounters the true Goal, “in the box.” Tanenhaus et al. demonstrated that by manipulating the referential context within the visual scene, they could completely abolish this garden path:

• One-Apple Context (Referentially Ambiguous): When the display contained a single apple sitting on a towel, and an empty towel elsewhere on the table, listeners showed classic garden path effects. Within 200 to 300 milliseconds of hearing “on the towel”, their eyes launched anticipatory saccades to the empty towel, misinterpreting it as the destination. Upon hearing “in the box”, their gaze exhibited confusion, long fixation latencies, and corrective saccades back to the target apple.
• Two-Apple Context (Referentially Constrained): When the visual display contained two apples—one resting on a towel and one resting on a napkin—the garden path was utterly eliminated. Because linguistic communication adheres to pragmatic principles of referential efficiency, having two identical candidate referents forces the comprehender to expect a restrictive modifier specifying which apple is being discussed. Driven by the visual context, listeners immediately interpreted “on the towel” as a modifier (Low NP-attachment). Their eyes moved directly and exclusively to the apple on the towel, never once looking at the empty towel, and smoothly transitioned directly to the box upon hearing “in the box.”

The Visual World Paradigm proved that sentence comprehension is fundamentally situated and interactive. Extra-linguistic visual context, pragmatic referential uniqueness constraints, and spoken phonological input are integrated simultaneously in real time, with ocular saccades responding to referential constraints within a mere 250 milliseconds. This empirical discovery profoundly undermined the strict modular claim that initial parsing passes are blind to non-syntactic information, establishing continuous constraint satisfaction as a dominant theoretical paradigm in modern cognitive science.

9. Cross-Linguistic Investigations of Garden Path Eye-Movement Patterns

9.1 Head-Final Languages: Parsing in German, Japanese, and Korean

The overwhelming majority of early psycholinguistic parsing models—including Bever’s canonical NVN strategy and Frazier’s Garden-Path Model—were formulated based primarily on empirical investigations of the English language. English is a structurally constrained, predominantly right-branching, Subject-Verb-Object (SVO) language characterized by relatively impoverished overt inflectional morphology and strict positional word order constraints. To test whether parsing heuristics reflect universal biological architecture or language-specific typographic adaptations, experimental researchers expanded eye-tracking investigations into Head-Final (SOV) languages such as German, Japanese, and Korean.

Head-final languages present a profound theoretical challenge to standard parsing models:

• Delayed Verbal Heads: In typical subordinate clauses in German, and throughout matrix clauses in Japanese and Korean, the verbal predicate that assigns thematic roles and subcategorizes clausal arguments appears at the absolute end of the clause. Consequently, the parser cannot utilize verb-driven argument structure biases or head-driven projection to build initial syntactic trees.
• Incremental Parsing Without Heads: Eye-tracking experiments in Japanese and German demonstrate that comprehenders do not maintain a passive, uncommitted holding buffer until the clause-final verb appears. Instead, readers actively build incremental syntactic representations upon encountering initial Case-marked nominals. In German verb-final subordinate clauses (e.g., “…dass die Mutter die Tochter…”), where the surface morphology is ambiguous between nominative and accusative case, readers systematically fall into garden paths if the clause-final verb requires an unexpected case-marking or thematic frame.
• Oculomotor Disambiguation Signatures: Eye-tracking records in German reveal massive inflations in first-pass gaze durations and regressions that are tightly localized to the clause-final verb. Because the structural ambiguity is maintained over a substantially longer spatial and temporal distance than in English, the reanalysis overhead in head-final languages often manifests as prolonged, multi-hop regressive saccades that traverse back across multiple intervening arguments, demonstrating that incremental structure-building is a universal imperative of the human parser, even in typographic environments devoid of immediate head information.

9.2 Relative Clause Attachment Ambiguities Across Languages

The universality of Frazier and Rayner’s Late Closure principle was subjected to severe empirical scrutiny following the groundbreaking cross-linguistic work of Fernando Cuetos and Don C. Mitchell in 1988. Cuetos and Mitchell investigated Relative Clause (RC) attachment ambiguities in complex nominal phrases involving two potential noun phrase hosts separated by a genitive preposition:

“The journalist interviewed the daughter of the colonel who had the accident.”

In this construction, the relative clause “who had the accident” can attach either to the second, most recently processed noun phrase (Low Attachment / NP2: “the colonel”) or to the first, higher noun phrase (High Attachment / NP1: “the daughter”).

The cross-linguistic findings produced an intense theoretical debate:

• English Preferences: English readers demonstrate a consistent, robust behavioral preference for Low Attachment (NP2), fully aligning with the Principle of Late Closure (attach incoming items to the constituent currently under active processing).
• Spanish, French, and Dutch Preferences: Cuetos and Mitchell demonstrated that native speakers of Spanish reading equivalent constructions (“la hija del coronel que…”) exhibit a statistically significant, baseline preference for High Attachment (NP1). Subsequent cross-linguistic eye-tracking experiments confirmed that French, Dutch, German, and Italian readers also show pronounced High Attachment preferences, systematically violating the Late Closure principle.

This cross-linguistic divergence forced a fundamental re-evaluation of parsing theory. Proponents of constraint-satisfaction proposed the Tuning Hypothesis (Mitchell et al., 1995), which posits that the human parsing architecture is statistically tuned by exposure: the parser’s initial structural attachments simply reflect the relative frequency with which specific structural configurations occur in natural discourse within that specific linguistic community. Simultaneously, prosodic theorists demonstrated that cross-linguistic variations in implicit prosodic phrasing and the presence of grammatical gender agreement markers (which are rich in Spanish and impoverished in English) heavily modulate ocular fixation distributions, disproving the notion of an invariant, hard-wired structural preference across all human languages.

9.3 Morphologically Rich Languages and Morphosyntactic Disambiguation

In morphologically rich languages—such as Russian, Polish, Finnish, and modern Slavic tongues—grammatical relationships are signaled primarily through elaborate systems of overt nominal case inflection, verbal agreement, and declension paradigms, rather than rigid positional word order. In these languages, word order is extraordinarily flexible; sentences can fluidly adopt SVO, OVS, SOV, or VSO configurations to reflect pragmatic topic-focus dynamics without altering propositional truth values.

Eye-tracking investigations in these typographic environments illuminate how the human parser adjudicates conflicts between word-order heuristics and morphological case cues:

• Morphological Precedence: When an unambiguous morphological case marker (e.g., an explicit accusative feminine suffix) appears on the sentence-initial noun, the parser instantly suppresses Bever’s NVN agent-first heuristic. Eye-tracking data show that Slavic readers immediately interpret an initial accusative nominal as the Patient/Theme, projecting an OVS or OSV structure without incurring processing slowdowns or elevated first-pass reading times.
• Morphological Ambiguity and Syncretism: However, when morphological case markers are syncretic (where a single surface inflectional ending is ambiguous between Nominative, Accusative, or Genitive case), the parser instantly defaults to Bever’s canonical Subject-first heuristic. If downstream disambiguating case markers (or verbal agreement morphology) subsequently reveal that the initial ambiguous nominal was actually an object, readers experience violent garden path disruptions.
• Oculomotor Recovery Trajectories: The eye-movement record in morphologically rich languages demonstrates that reanalysis is exceptionally swift when driven by explicit morphological agreement mismatch. The gaze executes rapid, microscopic regressions targeting the specific inflectional suffixes of preceding words, indicating that the cognitive repair mechanism leverages morphological features as precise computational coordinates to restructure the internal syntactic hierarchy.

10. Neurocognitive Correlates and Complementary Methodologies

10.1 Co-registration of Eye-Tracking and Electroencephalography (Eye-Fixation-Related Potentials)

For decades, psycholinguistics operated with an unfortunate methodological division: researchers utilized either high-precision eye-tracking (which provides exceptional ecological validity and natural ocular movement dynamics at the expense of indirect neural indexing) or high-temporal-resolution Event-Related Potentials (ERPs) derived from electroencephalography (EEG). Traditional ERP paradigms required unnatural Rapid Serial Visual Presentation (RSVP), wherein words are flashed centrally on a screen one at a time at fixed 300–500 millisecond intervals while participants suppress all blinks and ocular saccades to prevent electrical artifacts. This artificial paradigm destroyed natural reading velocity, prevented parafoveal preview, and made regressive reanalysis impossible.

In recent years, the methodological frontier has been fundamentally transformed by the development of co-registration methodologies: the simultaneous, continuous recording of eye movements and EEG, isolating Eye-Fixation-Related Potentials (EFRPs) (also known as Fixation-Related Potentials, or FRPs). In an EFRP experiment, the reader scans natural, continuous lines of text completely unrestrained. High-speed eye-trackers record every fixation coordinate and synchronize temporal triggers with the ongoing continuous EEG trace.

Advanced computational algorithms—such as independent component analysis (ICA) and continuous deconvolution modeling—are deployed to mathematically isolate and subtract the massive electrical artifacts produced by corneal-retinal dipoles during saccadic rotations. The resulting EFRP trace provides a millisecond-by-millisecond electrophysiological readout of the brain’s internal activity time-locked to the precise millisecond an ocular fixation lands on a critical target word. EFRP studies on garden path processing have confirmed that the electrophysiological markers of syntactic failure—traditionally observed only under artificial RSVP conditions—occur with stunning temporal fidelity during natural ocular reading, directly co-registering with first-pass fixations and the initiation of regressive saccades.

10.2 The P600 Component and the Syntactic Positive Shift

The electrophysiological signature most famously and robustly associated with garden path sentences is the P600 component (also termed the Syntactic Positive Shift, or SPS), originally discovered in landmark studies by Lee Osterhout and Phillip Holcomb (1992), and independently by Peter Hagoort and colleagues (1993). The P600 is a slow, positive deflection in the event-related electrical potential that typically onsets around 500 to 600 milliseconds following the presentation of a target word, peaking between 600 and 900 milliseconds, with a characteristic centroparietal scalp distribution.

The neurocognitive implications of the P600 in garden path processing are profound:

• Syntactic Violation vs. Syntactic Reanalysis: While severe, ungrammatical word-category violations (e.g., *”The man walked the into room”*) frequently elicit an Early Left Anterior Negativity (ELAN) or a standard Left Anterior Negativity (LAN) peaking around 300–400 milliseconds, fully grammatical garden path sentences (like Bever’s “The horse raced past the barn fell”) typically bypass the early LAN and produce an enormous, prominent P600 at the disambiguating verb.
• Oculomotor-ERP Correspondence: Co-registration EFRP experiments have demonstrated a direct mathematical correspondence between the amplitude of the P600 wave and the spatial duration of late-stage eye-tracking metrics. The greater the amplitude of the parietal P600 elicited by the disambiguating token, the longer the Regression Path Duration (Go-Past Time) and the higher the probability of executing an immediate regressive saccade.
• Cognitive Meaning of the P600: Theoretical consensus interprets the P600 not as an index of initial syntactic tree construction, but as a direct neurocomputational reflection of syntactic reanalysis, structural repair, and secondary integration. It represents the intensive neural effort deployed by distributed frontoparietal networks to diagnose structural incompatibility, suppress the incorrect canonical parse, and rebuild the phrase marker to accommodate the licensed syntactic structure.

10.3 Hemodynamic Correlates: fMRI and Functional Neuroanatomy of Parsing

While eye-tracking and ERPs provide millisecond-level temporal resolution, functional Magnetic Resonance Imaging (fMRI) provides spatial localization, mapping the precise neuroanatomical circuits responsible for parsing garden path sentences. Hemodynamic investigations have mapped the distributed neurocognitive network that mobilizes when comprehenders encounter syntactic ambiguity and processing breakdown.

The core neural substrate of garden path resolution comprises:

• Left Inferior Frontal Gyrus (LIFG / Broca’s Area): Hemodynamic activation during garden path processing is heavily concentrated in the left inferior frontal gyrus, specifically encompassing Brodmann Area 44 (pars opercularis) and Brodmann Area 45 (pars triangularis). Advanced neuroimaging studies consistently demonstrate that while simple canonical sentences generate modest baseline activation in Broca’s area, the introduction of non-canonical reduced relative clauses or sentential complement ambiguities triggers massive signal increases in BA 44/45. Crucially, functional dissociations suggest that BA 44 is primarily engaged in the mechanical restructuring of hierarchical phrase trees, whereas BA 45 coordinates cognitive control, working memory retrieval, and the suppression of the competing, incorrect semantic parse.
• Left Posterior Superior Temporal Gyrus (pSTG / Wernicke’s Area): Robust activations are concurrently observed in the left posterior superior temporal gyrus and the superior temporal sulcus. This region is intimately involved in lexical-syntactic retrieval, specifically accessing and mapping the complex subcategorization frames and argument structures of ambiguous verbs.
• Frontoparietal Cognitive Control Network: During severe garden path reanalysis, the language-specific network recruits domain-general cognitive control hubs, including the dorsal Anterior Cingulate Cortex (dACC) and bilateral frontoparietal networks. The dACC activates to monitor computational conflict and flag the categorical error, while the frontoparietal network manages the sustained working memory load required while the gaze traverses backward across upstream text to reassemble the linguistic representation.

11. Contemporary Re-evaluations: Good-Enough Processing and Surprisal Theory

11.1 Ferreira’s Good-Enough Processing Framework

For decades, virtually all psycholinguistic parsing theories—whether modular like the Garden-Path Model or interactive like Constraint-Satisfaction—shared an uncontested foundational assumption: that the ultimate objective of the human parsing mechanism is to construct a fully detailed, mathematically complete, and globally veridical syntactic tree of every sentence it encounters. In a radical and influential re-evaluation, Fernanda Ferreira and her collaborators formulated the Good-Enough Processing Framework, demonstrating that the human mind routinely settles for shallow, incomplete, and systematically inaccurate linguistic representations.

Ferreira (2003) conducted a series of seminal experiments probing readers’ final semantic interpretations of classic garden path sentences. After participants read sentences such as “While the man hunted the deer ran into the woods,” they were presented with direct, binary comprehension questions, such as: “Did the man hunt the deer?”

• The Persistence of the Garden Path: According to standard theories, if a reader successfully reanalyzes the sentence (progressing past the disambiguating verb “ran” and reading fluently to the end), the incorrect initial direct object parse should be purged from memory, leaving only the correct representation: the man was hunting unspecified game, and the deer was the entity running. Astonishingly, Ferreira found that an overwhelming majority of participants answered “Yes”—affirming that the man hunted the deer.
• Shallow Oculomotor Passes: Complementary eye-tracking experiments revealed that when readers execute regressive saccades in garden path sentences, their ocular re-reading passes are frequently superficial. Readers often glance back just long enough to resolve the localized grammatical blockage at the disambiguating verb, without executing the exhaustive structural rebuild necessary to completely decouple the initial thematic bindings.
• Dual-Track Mental Representations: Ferreira demonstrated that human language processing operates via two parallel tracks: a heuristic, shallow “fast” processor that relies on Bever’s canonical templates (yielding the persistent belief that the man hunted the deer), and an algorithmic, slower syntactic parser. If the heuristic output satisfies general communicative plausibility, the brain frequently accepts a “good-enough” representation, allowing persistent misinterpretations of garden path sentences to endure indefinitely in long-term episodic memory.

11.2 Information-Theoretic Surprisal and Entropy Reduction

In contemporary computational psycholinguistics, structural parsing has been mathematically revolutionized by the integration of Information Theory, primarily through the framework of Surprisal Theory, formulated by John Hale (2001) and fully articulated by Roger Levy (2008). Surprisal theory bridges the gap between formal algorithmic parsing, probabilistic generative grammars, and empirical ocular fixations, providing an overarching mathematical unification of Bever’s original cognitive strategies.

Surprisal theory formalizes the cognitive processing difficulty of a word as its negative log-probability of occurrence, conditioned on the preceding linguistic context:

Surprisal(w_i) = -log₂ P(w_i | w₁, w₂, …, w_i-1)

Under this computational framework:

• Mathematical Definition of the Garden Path: In an ambiguous sentence like “The horse raced past the barn fell,” the prefix “The horse raced past the barn…” assigns an overwhelmingly high conditional probability to continuing structures that complete an active matrix clause (or terminate the sentence entirely). The conditional probability of encountering an intransitive finite matrix verb like “fell” is extraordinarily near zero: P(“fell” | Context) → 0. Consequently, the information-theoretic Surprisal value at “fell” approaches infinity: -log₂(0) → ∞.
• Linear Predictor of Fixation Durations: Surprisal theory posits an explicit, lawful relationship: the cognitive processing effort—and consequently, the First-Pass Reading Time and Gaze Duration observed on an eye-tracker—scales linearly with the surprisal value of the word. Computational models utilizing probabilistic context-free grammars (PCFGs) and n-gram language models successfully simulate the exact millisecond-level fixation elevations observed at disambiguating regions in human experimental subjects.
• Entropy Reduction: Complementary work by Hale formalizes garden path reanalysis as Entropy Reduction: the abrupt, massive collapse of an uncertain probability distribution over multiple possible syntactic trees into a single, highly constrained, and structurally complex parse. Surprisal theory mathematically legitimizes Bever’s 1970 insights, demonstrating that perceptual heuristics are optimal statistical approximations of information density across natural language environments.

11.3 Neural Network and Large Language Model Parsing Dynamics

The meteoric rise of deep transformer-based Large Language Models (LLMs)—such as GPT-4, LLaMA, and BERT—has introduced a novel computational testing ground for sentence processing research. Psycholinguists and computational neuroscientists now utilize modern transformer networks as “silicon participants,” testing whether deep neural architectures trained purely on statistical text prediction exhibit internal representations and processing breakdowns analogous to human garden path dynamics.

Recent research analyzing the attention-weight distributions and internal layer states of transformer architectures reveals striking parallels to human parsing behavior:

• Attention-Weight Reorganization: When a transformer processes “The horse raced past the barn fell,” the self-attention heads in the initial and intermediate layers overwhelmingly direct attention from “raced” back to “The horse” as an Agent-Verb dependency. When the token “fell” is ingested at the input layer, higher transformer layers exhibit massive internal entropy and conflict. The attention weights undergo radical, violent reallocation: attention heads must retroactively suppress the strong dependency link between “horse” and “raced” and project a direct long-distance dependency connecting “horse” to “fell.”
• Model Surprisal vs. Human Eye Fixations: By extracting token-level surprisal scores directly from the output softmax distributions of LLMs, researchers can directly predict human reading times recorded via eye-tracking. Modern transformer surprisal metrics achieve unprecedented accuracy in predicting human Gaze Durations, Regression Path Durations, and Skipping Rates across classic garden path benchmarks, demonstrating that human-like parsing heuristics naturally emerge from deep statistical learning across massive corpora.
• Silicon vs. Biological Divergences: Despite these remarkable convergences, significant differences remain. Standard transformer models process context bidirectionally or retain all prior token representations perfectly within dense vector memory across unlimited temporal spans without decay. Unlike the biological human parser, transformer models do not possess biological working memory constraints, metabolic energetic costs, or the physical necessity of ocular saccades. Studying where modern LLMs successfully parse garden paths that systematically trip up human readers continues to delineate the exact biological and ocular constraints that define human cognitive architecture.

12. Methodological Implications and the Future of Eye-Tracking in Parsing Research

12.1 Statistical Advances: Linear Mixed-Effects Modeling in Reading Data

Alongside theoretical revolutions, the empirical analysis of eye-tracking reading data has undergone an absolute statistical transformation. Historically, psycholinguists analyzed eye-tracking duration metrics using standard repeated-measures Analysis of Variance (ANOVA), necessitating dual separate analyses: one averaging across participants (F1 analysis) and one averaging across experimental sentence items (F2 analysis). This classic Clark (1973) “language-as-fixed-effect fallacy” approach was methodologically flawed: it required discarding massive amounts of trial-by-trial variance through aggregation, struggled with missing data points (ubiquitous in eye-tracking when words are skipped), and failed to simultaneously model crossed random variations of both human subjects and linguistic stimuli.

In modern psycholinguistics, standard practice mandates the deployment of Linear Mixed-Effects Models (LMMs) and Generalized Linear Mixed Models (GLMMs), typically implemented in statistical computing environments such as R (via packages like lme4 and brms):

• Simultaneous Crossed Random Effects: LMMs permit the simultaneous inclusion of crossed random intercepts and random slopes for both participants and items within a single unified statistical model, preventing type I error inflation while fully preserving trial-level variance.
• Appropriate Error Distributions: Because eye-fixation durations are intrinsically non-normal, continuous, positively skewed variables (characterized by long right-hand tails), modern statistical pipelines model raw durations using generalized families, such as Gamma or log-normal distributions, rather than forcing inappropriate Gaussian assumptions onto millisecond data.
• Binary Logistic Modeling for Regressions: For discrete, categorical eye-movement metrics—such as the probability of executing a Regression Out or skipping a word—GLMMs utilizing logit link functions provide robust statistical power to determine whether structural garden paths induce statistically significant spikes in backward saccades while controlling for word length, frequency, and trial order effects.

12.2 Individual Differences: Age, Working Memory, and Literacy Variations

Contemporary sentence processing research has increasingly abandoned the fictional abstraction of the “ideal, universal comprehender,” turning the ocular lens toward individual differences across diverse human populations. High-precision eye-tracking has demonstrated that the oculomotor manifestation of garden path reanalysis is not monolithic, but is profoundly shaped by an individual’s cognitive architecture, developmental stage, and life experience.

Key dimensions of individual variation include:

• Working Memory Capacity (Reading Span): As pioneered by Just and Carpenter, readers with high working memory capacity demonstrate significantly more efficient garden path recovery. Their eye-movement records show targeted, precise regressive saccades that land directly on the ambiguous head, coupled with swift resolution in Go-Past Time. Conversely, readers with low working memory capacity frequently exhibit scattered, disorganized regression paths, re-reading entire preceding sentences multiple times or abandoning the reanalysis pass entirely.
• Developmental Populations and Second-Language (L2) Learners: Developing children show prolonged first-pass fixations and an extreme over-reliance on Bever’s canonical NVN heuristic, often failing to execute corrective regressions when disambiguating morphology appears. Non-native (L2) readers exhibit unique parsing signatures: even when highly proficient, L2 readers frequently experience severe spillover effects, with garden path reanalysis costs lingering across several words downstream of the disambiguating region due to elevated lexical retrieval latencies.
• Aging and Compensatory Processing: Older adults (typically aged 65 and above) show distinct ocular adjustments during sentence processing. Due to reductions in working memory capacity and visual acuity, older readers exhibit higher baseline skipping rates for short words, longer first-pass fixation durations, and a reliance on “risky” predictive parsing. When tripped by a garden path sentence, older comprehenders display pronounced elevations in Second-Pass Reading Time and rely heavily on good-enough processing heuristics to maintain fluent reading velocity.

12.3 Future Paradigms: Virtual Reality, Mobile Eye-Tracking, and Naturalistic Reading

As eye-tracking technology continues its inexorable trajectory toward miniaturization and ubiquitous computing, the paradigms undergirding sentence processing experiments are expanding far beyond the sterile confines of chin-rest-bound desktop monitors displaying isolated, static sentences.

Emerging frontiers include:

• Naturalistic Continuous Discourse: Rather than evaluating artificially constructed, isolated sentences like “The horse raced past the barn fell,” contemporary researchers are deploying eye-tracking across entire naturalistic novels, magazine articles, and multi-paragraph narratives. The construction of massive, open-access natural reading corpora (such as the Dundee Corpus and the Provo Corpus) allows psycholinguists to evaluate garden path dynamics and surprisal predictions across thousands of unconstrained, ecologically authentic contexts.
• Immersive Virtual Reality (VR) and Mobile Eye-Tracking: The integration of high-speed infrared eye-trackers directly into Virtual Reality and Augmented Reality headsets allows researchers to merge the Visual World Paradigm with dynamic, fully embodied, three-dimensional physical environments. Comprehenders can navigate realistic virtual worlds, manipulating objects while processing complex, structurally ambiguous auditory commands, providing unprecedented ecological fidelity for studying situated, multimodal parsing.
• Crowdsourced Webcam Eye-Tracking: While historically limited by poor spatial resolution, cutting-edge machine-learning algorithms and deep neural networks are increasingly capable of tracking ocular gaze using standard consumer webcams. While webcam tracking cannot yet rival the sub-millisecond precision of dedicated 2000-Hz laboratory systems for measuring subtle First Fixation Durations, it provides unprecedented scalability, enabling the collection of structural parsing and reading data from tens of thousands of participants across global, linguistically diverse populations simultaneously.

Conclusion

The journey from Thomas G. Bever’s intuitive 1970 perceptual strategies to the sophisticated, multi-modal psycholinguistic laboratories of the present day represents one of the most intellectually fruitful arcs in the history of cognitive science. When Bever first scribbled down the deceptively unassuming sentence “The horse raced past the barn fell,” he provided the nascent cognitive revolution with a profound methodological realization: that the latent, invisible architecture of the human mind is most brilliantly illuminated precisely at the moment it breaks down. The systemic failure of the human sentence processor to anticipate the true structural trajectory of a reduced relative clause exposed the internal operating system of real-time human cognition—revealing an incremental, highly heuristic computational engine ruthlessly optimized for temporal speed, cognitive economy, and instantaneous interpretation.

The subsequent empirical marriage between Bever’s foundational insights and high-precision eye-tracking methodology transformed speculative theoretical models into observable, quantifiable behavioral physics. By mapping the microsecond-level mechanics of ocular fixations, gaze durations, saccadic leaps, and regressive trajectories, researchers like Lyn Frazier, Keith Rayner, Marcel Just, and their contemporaries provided an unmediated window into the instantaneous operations of the parsing mind. The ocular lens decisively answered questions that formal generative linguistic competence could never resolve alone: proving that parsing commitments are made incrementally without delay, detailing the steep energetic tax of structural reanalysis, and charting the fine-grained computational competitions between encapsulated geometric rules and rich, real-time contextual constraints.

Today, as sentence processing research continues to evolve—integrating information-theoretic surprisal models, co-registered neurophysiological recordings, good-enough shallow heuristics, and deep neural network simulations—the legacy of the classic garden path experiments remains undiminished. The simple observation that human beings can be reliably led down a structural blind alley continues to serve as an indispensable paradigm for testing theories of human language, memory, and cognitive architecture. In an era where artificial intelligence systems process linguistic strings through billions of uninterpretable silicon parameters, the study of human eye-tracking during garden path sentence processing stands as an enduring testament to the power of experimental cognitive psychology: demonstrating that by meticulously observing the physical movements of the human eye, we can systematically decipher the magnificent, biological computational mysteries of the human mind.

References

• Bever, T. G. (1970). The cognitive basis for linguistic structures. In J. R. Hayes (Ed.), Cognition and the development of language (pp. 279–362). John Wiley & Sons.
• Carpenter, P. A., & Just, M. A. (1983). What your eyes do while your mind is reading. In K. Rayner (Ed.), Eye movements in reading: Perceptual and language processes (pp. 275–307). Academic Press. https://doi.org/10.1016/B978-0-12-583680-7.50020-0
• Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press. https://mitpress.mit.edu/9780262527408/aspects-of-the-theory-of-syntax/
• Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the use of the Late Closure strategy in Spanish. Cognition, 30(1), 73–105. https://doi.org/10.1016/0010-0277(88)90004-2
• Ferreira, F. (2003). The misinterpretation of noncanonical sentences. Cognitive Psychology, 47(2), 164–203. https://doi.org/10.1016/S0010-0285(03)00005-7
• Ferreira, F., Bailey, K. G., & Ferraro, V. (2002). Good-enough representations in language comprehension. Current Directions in Psychological Science, 11(1), 11–15. https://doi.org/10.1111/1467-8721.00158
• Fodor, J. A. (1983). The Modularity of Mind: An Essay on Faculty Psychology. MIT Press. https://mitpress.mit.edu/9780262560252/the-modularity-of-mind/
• Fodor, J. D., & Inoue, A. (1994). The diagnosis and cure of garden paths. Journal of Psycholinguistic Research, 23(5), 407–434. https://doi.org/10.1007/BF02143940
• Frazier, L. (1979). On comprehending sentences: Syntactic parsing strategies [Doctoral dissertation, University of Connecticut]. Indiana University Linguistics Club.
• Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14(2), 178–210. https://doi.org/10.1016/0010-0285(82)90008-1
• Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997). The contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of Memory and Language, 37(1), 58–93. https://doi.org/10.1006/jmla.1997.2512
• Hagoort, P., Brown, C., & Groothusen, J. (1993). The Syntactic Positive Shift (SPS) as an ERP measure of syntactic processing. Language and Cognitive Processes, 8(4), 439–483. https://doi.org/10.1080/01690969308407585
• Hale, J. (2001). A probabilistic Earley parser as a psycholinguistic model. In Proceedings of the second meeting of the North American Chapter of the Association for Computational Linguistics on Language technologies (pp. 1–8). Association for Computational Linguistics. https://doi.org/10.3115/1073336.1073357
• Just, M. A., & Carpenter, P. A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87(4), 329–354. https://doi.org/10.1037/0033-295X.87.4.329
• Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126–1177. https://doi.org/10.1016/j.cognition.2007.05.006
• MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101(4), 676–703. https://doi.org/10.1037/0033-295X.101.4.676
• Osterhout, L., & Holcomb, P. J. (1992). Event-related brain potentials elicited by syntactic anomaly. Journal of Memory and Language, 31(6), 785–806. https://doi.org/10.1016/0749-596X(92)90039-Z
• Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124(3), 372–422. https://doi.org/10.1037/0033-2909.124.3.372
• Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., & Sedivy, J. C. (1995). Integration of visual and linguistic information in spoken language comprehension. Science, 268(5217), 1632–1634. https://doi.org/10.1126/science.7777863
• Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33(3), 285–318. https://doi.org/10.1006/jmla.1994.1014