RESEARCH DOSSIER
Sentence Processing and Syntactic Parsing
Abstract. Sentence processing combines incremental structural analysis with lexical, semantic, contextual, and memory-dependent operations. This fundamental research monograph critically synthesizes experimental and computational findings on syntactic parsing: initial structural commitment, garden-path disruption, persistent misinterpretations, attachment decisions, filler–gap dependencies, dependency locality, retrieval interference, agreement attraction, and probabilistic expectations. It examines the methodological strengths and constraints of self-paced reading, eye tracking, visual-world paradigms, event-related potentials, functional neuroimaging, and cross-language comparisons. Serial, constraint-based, expectation-based, locality-based, and cue-based retrieval accounts are evaluated against original studies, replication research, and contested findings on Chinese relative clauses. A dedicated contemporary section integrates 2023–2025 results on cross-linguistic surprisal, syntactic disambiguation benchmarks, limits of model scaling, and Transformer accounts of retrieval and prediction. The synthesis specifies evidence-supported conclusions, unresolved theoretical distinctions, and a comparative research agenda for human and artificial language-processing systems. This work is a critical narrative synthesis of published research; it reports no new participants or quantitative meta-analysis.
Keywords: sentence processing; syntactic parsing; garden-path sentences; incremental comprehension; syntactic ambiguity; syntactic reanalysis; surprisal; dependency locality; cue-based retrieval; agreement attraction; filler–gap dependencies; eye tracking; self-paced reading; event-related potentials; psycholinguistics
Sentence Comprehension as an Incremental Computation
Sentence processing is the sequence of perceptual, lexical, grammatical, and interpretive operations by which a listener or reader incrementally constructs relations among words. Syntactic parsing is its structural component: incoming expressions are assigned categories, grouped into constituents, and linked by dependencies that constrain the roles of participants and the interpretation of events. The scientific object of this dossier is the time course of those assignments, their revision, and cases in which revision remains incomplete. An adequate model predicts what a comprehender considers at a particular word, how much processing it costs, and what interpretation survives when the sentence ends. The eye-movement experiments of Frazier & Rayner (1982) and probabilistic formalization by Levy (2008) make these questions experimentally and computationally precise.
Consider the locally ambiguous sequence 'While the researcher interviewed the assistant ...'. Before the next clause begins, 'the assistant' may be taken as the interviewee. A continuation such as 'the assistant took notes' instead makes the noun phrase the subject of a new clause, requiring a different structural assignment. The garden-path effect refers to processing disruption when the favored analysis becomes untenable. It is diagnosed by measured behavior, including increased fixation durations, regressions, slowed self-paced reading, or comprehension errors; no single example proves that all readers compute exactly one initial tree. Frazier & Rayner (1982) documented early disambiguation costs, while Christianson et al. (2001) showed that an earlier thematic interpretation can persist despite subsequent structural repair.
Parsing is partly determined by the grammar of a language and partly by the processing architecture of its users. Word order, agreement, case, lexical biases, discourse referents, attention, and working-memory operations can all alter the trajectory of comprehension. The scientific challenge is to specify the contribution of each cue, rather than to infer hidden cognitive stages directly from one aggregate reading-time difference. This monograph is a critical narrative synthesis of original experiments, formal models, neuroscientific studies, and cross-language evidence. It does not report new participants, a registered systematic review, or a quantitative meta-analysis. Its comparisons distinguish measurable phenomena from explanatory commitments, drawing especially on MacDonald et al. (1994), Lewis et al. (2006), and Vasishth, von der Malsburg, & Engelmann (2013).
Constructs, Representations, and Explanatory Levels
A parser may assign phrase structure, identify hierarchical relations, attach modifiers, establish dependencies between separated expressions, and connect words to thematic roles. These operations are related but analytically separable. Attachment asks which constituent an incoming phrase modifies; filler-gap resolution asks which unexpressed argument position is linked to an earlier phrase; thematic assignment asks who performed an action and who underwent it. The same overt sentence can support several representations during its unfolding. The experimental outcome called a garden path may reflect initial attachment preferences, retrieval failure, reinterpretation of thematic roles, or a combination of these processes. Trueswell et al. (1994) demonstrates why semantic constraints must be modeled explicitly when interpreting structural ambiguity.
An explanatory theory needs a representation space and a processing procedure. A grammar specifies admissible relations, but grammar alone ordinarily leaves unresolved which structure is entertained first, which information changes its probability, and how incomplete hypotheses are stored. A processing model adds commitments about timing, resource allocation, competition, and repair. Constraint-based models assign graded support to candidate analyses MacDonald et al. (1994); locality models link difficulty to structural distance Gibson (1998); expectation-based models assign cost to improbable continuations Levy (2008). Their differences can be examined at the precise regions of a sentence where predictions diverge. A formal model's apparent success depends on both its grammatical inventory and the empirical estimate of the cues it uses.
Sentence-level meaning is a further explanatory level. A reader may generate a semantically plausible interpretation that conflicts with a fully specified grammatical analysis, or may preserve features of an abandoned interpretation. Ferreira et al. (2002) calls attention to such good-enough representations, and Christianson et al. (2001) supplies experimental evidence that residual thematic roles can linger. This is a substantive result because successful word recognition and even a partially successful parse do not guarantee accurate answers to questions about who did what. Studies should measure structural decisions and ultimate interpretation independently, using comprehension probes that do not introduce their own ambiguities. A complete account of processing needs both an online time course and an endpoint representation.
Historical Development of Experimental Parsing Research
Experimental sentence-processing research developed at the intersection of linguistic theory, information-processing psychology, and increasingly precise measures of behavior. Earlier linguistic proposals supplied formal distinctions between constituent structures and dependencies; the psycholinguistic program asked how people construct those structures under real-time constraints. In the 1980s, eye-tracking provided evidence that a structural preference can influence the very first encounter with disambiguating material. Frazier & Rayner (1982) recorded eye movements during the reading of temporarily ambiguous sentences and interpreted the pattern in terms of minimal attachment, late closure, and selective reanalysis. The decisive methodological contribution was to locate disruption within an unfolding sentence rather than infer an entire parsing sequence from a final judgment alone.
The modularity debate was sharpened by Ferreira & Clifton (1986), who examined whether semantic plausibility and discourse context eliminate early syntactic preferences. Their experiments included both eye-movement and self-paced-reading measurements, and their interpretation favored a relatively independent early syntactic operation. Trueswell et al. (1994) subsequently revisited reduced-relative ambiguities with tighter control of animacy, thematic fit, and disambiguating regions; they reported immediate semantic influences that challenged the inference of a universally encapsulated initial stage. These landmark studies remain important because they show how seemingly abstract architectural claims turn on lexical materials, baseline conditions, and the mapping between a processing effect and the time course of information use.
During the 1990s and 2000s, competing accounts became more explicit about distributional experience, memory, and computational complexity. MacDonald et al. (1994) integrated lexical ambiguity and structural competition within a constraint-based framework. Gibson (1998) formalized memory and integration demands associated with dependencies; Hale (2001) and Levy (2008) developed probabilistic predictions from structural expectations. Lewis et al. (2006) related syntactic retrieval to broader principles of working memory. This history is best read as progressive expansion of explanatory targets: processing theories must now address immediate preferences, contextual modulation, reanalysis, graded uncertainty, resource limitations, and the final interpretation, sometimes in different languages and modalities. Historical priority for a theory does not establish universal empirical validity.
Incrementality and Structural Commitment
Incremental processing means that comprehenders use an unfolding prefix of the input before the entire sentence is available. A lexical item can narrow possible attachments, open a syntactic expectation, or trigger integration with a dependent encountered earlier. Some theories posit that the parser rapidly selects a preferred analysis, while others allow graded competition among analyses. The distinction matters especially when early information favors one structure but the next phrase requires another. Frazier & Rayner (1982) identifies the temporal signature of such conflict in reading; Hale (2001) shows that probabilistic parsers can make explicit predictions about costs at successive words. Neither behavioral slowing nor an expectation model by itself reveals the complete contents of working memory.
Structural commitment has several meanings that should remain distinct. It may describe a temporary decision to attach an incoming phrase, a sharply skewed probability distribution over alternatives, or an interpretation so strong that later evidence cannot fully undo it. A parallel model can behave as though it were serial when one candidate dominates the distribution; a serial model can recover quickly when disambiguation occurs early. This explains why the existence of garden-path costs does not by itself discriminate uniquely between serial and parallel processing. A stronger experiment manipulates prior probabilities, conflicting constraints, and the exact location of disambiguation, then compares the complete predicted time courses MacDonald et al. (1994) Levy (2008).
Immediate processing also depends on the scale of the evidence. An eye fixation reflects visual sampling and oculomotor decisions as well as linguistic processing; a self-paced reading keypress reflects a decision to proceed. Vasishth, von der Malsburg, & Engelmann (2013) emphasizes that comprehension and eye-movement control interact in ways that complicate direct temporal inference. Late effects may arise from an earlier processing event propagating across word boundaries, and early-looking effects can vary with how regions are segmented. Researchers should therefore report first-pass and rereading measures, downstream spillover regions, excluded trials, and the instructions given to participants. Converging measures improve the temporal interpretation of structural commitment.
Garden Paths, Reanalysis, and Persistent Interpretations
A garden-path sentence is a controlled probe of how people revise an initially compelling syntactic hypothesis. The relevant scientific comparison pairs a locally ambiguous input with a closely matched unambiguous counterpart and examines behavior at and after disambiguation. Frazier & Rayner (1982) found increased reading costs when disambiguation contradicted preferred attachment strategies, including effects in initial fixation behavior and selective eye regressions. These outcomes show that readers make decisions before reaching a sentence's end and subsequently adjust them. They do not establish a single universal repair algorithm: regressions can reflect attention shifts, targeted reinterpretation, or rereading that does not culminate in a fully corrected interpretation.
Repair may be structurally successful yet semantically partial. In the experiments of Christianson et al. (2001), participants could recognize the correct final event while retaining an incorrect relation that had been supported by the earlier garden-path analysis. Ferreira et al. (2001) reported analogous interpretive consequences of difficult reanalysis. This dissociation is central: a sentence can receive a globally grammatical parse while the reader's answer to a thematic question remains affected by an abandoned local assignment. Under a good-enough account Ferreira et al. (2002), comprehension is task-sensitive and may preserve incomplete or shallow interpretations when additional revision has little immediate benefit. The claim concerns experimentally observed accuracy under particular conditions, not an assertion that ordinary comprehension is uniformly defective.
Designing a valid garden-path study requires distinguishing syntactic revision from plausibility, memory, and probe effects. A comprehension question that restates the misleading relation can increase false endorsement; a paraphrase task may depend on recall after the sentence rather than online parsing. Ambiguity strength varies with lexical frequency and the semantic compatibility of an early noun phrase. The comparison between Ferreira & Clifton (1986) and Trueswell et al. (1994) illustrates why materials that superficially share a grammatical form may deliver substantially different early cues. Strong evidence therefore combines region-by-region processing data, unambiguous baselines, multiple items, and interpretive tests that identify exactly which earlier commitments survive.
Minimal Attachment, Late Closure, and Serial Accounts
Serial parsing accounts provide a particularly clear explanation of garden-path difficulty: the parser initially constructs one favored structural analysis and revises that analysis when subsequent input contradicts it. In classical garden-path theory, minimal attachment favors structures that add relatively little new syntactic machinery, while late closure favors linking incoming material to the clause or constituent currently being processed. Frazier & Rayner (1982) used reading-time and eye-movement patterns to test the consequences of such strategies. These principles are hypotheses about how limited cognitive resources are allocated, and they require operational interpretations of what constitutes a simpler analysis. A structural preference must be stated before examining the outcome of a particular item.
The explanatory value of a serial strategy lies in its specificity. If an ambiguous prepositional phrase can attach at two sites, a serial model predicts which site is assigned first; if a disambiguating word invalidates that attachment, it predicts the need for revision. Yet structural simplicity and recency can correlate with lexical probabilities and discourse expectations, producing similar behavioral predictions across architectures. Traxler et al. (1998) examined adjunct attachments in eye-tracking experiments and argued that the mechanism cannot be reduced to lexical ambiguity resolution alone. Such results caution against importing one parsing principle wholesale across attachment types. A general model must distinguish argument structures, optional modifiers, and constructions with different grammatical commitments.
Testing a serial account requires more than a contrast between ambiguous and unambiguous sentences. The design should include conditions in which two structural principles pull in different directions, with independent controls for lexical frequency, length, plausibility, and discourse context. Process-level predictions should specify whether a bias affects first-pass fixation, regression probability, or late comprehension accuracy. Ferreira & Clifton (1986) offered evidence for early syntactic biases under semantically disfavored conditions; Trueswell et al. (1994) found a different pattern when lexical-thematic cues were carefully adjusted. The empirical disagreement defines a useful research program: determine when early commitment remains robust and when contextual evidence changes the initial analysis itself.
Constraint-Based Competition and Lexical Information
Constraint-based theories treat sentence comprehension as competition among structural and interpretive alternatives under multiple graded sources of information. These include how frequently a verb occurs with a particular complement structure, the plausibility of candidate agents and patients, discourse referents, and the distribution of constructions in prior experience. MacDonald et al. (1994) framed ambiguity resolution as intimately connected to the lexical properties of individual words. The model explains why two sentences with similar formal ambiguity can yield different reading patterns if their verbs and nouns support competing analyses with different strengths. Its testable claim is that information traditionally labeled semantic or contextual can shape structural choices during the unfolding input.
The evidence for immediate constraint interaction is especially visible in Trueswell et al. (1994), which reexamined reduced relatives in terms of animacy and thematic fit. The authors compared ambiguous and morphologically unambiguous controls, adjusted lexical materials, and found that the apparent garden path depended on whether a noun phrase was a plausible agent of its following verb. This result does not show that every discourse cue affects every parsing decision at the earliest measurable moment. It establishes that the assumption of universal early semantic independence cannot be accepted without specifying the stimuli and measurement conditions. In spoken interaction, Tanenhaus et al. (1995) further demonstrated that visual referential information can affect real-time linguistic interpretation.
Constraint-based explanation is strongest when it quantifies candidate structures and derives item-specific predictions before observing reading times. Experimental norms for verb subcategorization, thematic plausibility, and contextual expectations must be estimated independently to avoid circular inference. Natural corpora can supply frequencies, but those frequencies are sensitive to genre, discourse, register, and the annotation scheme. Roland et al. (2007) compared grammatical structure distributions in corpus material, illustrating the problem of assuming that a general frequency estimate straightforwardly represents every comprehender's experience. A model that predicts every outcome by freely reweighting constraints has weak explanatory content; the central methodological requirement is generalization to new items and participants.
Probability, Surprisal, and Syntactic Expectations
Probabilistic accounts describe the changing distribution of potential continuations as a sentence unfolds. Under surprisal theory, processing difficulty at word w is related to the negative logarithm of the conditional probability of w given the preceding context. The formulation formalizes the intuition that a continuation is costly when it forces a sharp revision of expectations, even if the new structure is locally simple. Hale (2001) explored a probabilistic Earley parser as a psycholinguistic model; Levy (2008) developed a wider expectation-based account connecting probabilistic disambiguation to processing costs. The resulting explanatory variable can be estimated from structured grammar models, corpora, or other language models, with important differences between their assumptions.
Surprisal can explain why some locally long dependencies are relatively easy when a specific word becomes highly predictable. A strong syntactic expectation may offset integration difficulty, creating cases where locality and predictability make opposing predictions. Levy (2008) identifies such contrasts as particularly informative. Corpus-based eye-tracking research by Demberg & Keller (2008) tested dependency-locality and surprisal predictors on natural reading data; its findings showed that the theories can capture different components of variation, rather than allowing a simple universal victory for one account. Because predictability and word frequency covary, regression models must include suitable controls and should assess their incremental predictive contribution on held-out material.
Evidence for expectation requires precise distinctions. Contextual facilitation of an upcoming word may reflect distributed activation of alternatives, a narrowed probability distribution, or a specific preactivated form. Those mechanisms are not observationally equivalent. Staub (2015) reviews the link between lexical predictability and eye movements and emphasizes graded activation, while Frank et al. (2015) found an association between word surprisal and N400 amplitude in ERP data. These results inform, but do not settle, the place of hierarchical syntactic prediction within neural and behavioral processing. Predictability should be modeled as a continuously estimated variable with uncertainty, rather than presented as a directly observed mental forecast of a single word.
Dependency Locality and Integration Costs
Dependency-based theories focus on the burden of keeping an incomplete relation active and integrating constituents separated by intervening material. Gibson (1998) proposed a theory combining integration cost with memory cost for unresolved predictions, and used it to discuss center embedding, relative clauses, and other syntactic complexity phenomena. The intuitive mechanism is that building a relation between distant expressions requires greater resources when intervening structure imposes competing representations or storage demands. The theoretical definition of distance matters: counting words, discourse referents, syntactic heads, or expected categories yields different predictions. A resource theory acquires empirical precision when it specifies the relevant unit and the location of increased cost.
Comparative testing reveals that dependency distance is an important candidate mechanism but is not a universal sufficient explanation. Demberg & Keller (2008) used reading-time data from the Dundee eye-tracking corpus and reported a mixed pattern: integration cost was informative for particular word classes, whereas a global effect over arbitrary words was not established. Naturalistic text offers ecological breadth but produces correlated predictors and heterogeneous discourse structures. Controlled experiments offer stronger manipulation of dependency length but sometimes use unusual sentences that alter participants' strategies. Both approaches are necessary to separate memory demands from probabilistic anticipation, lexical constraints, and semantic interpretability.
Locality can also change its apparent effect across languages with different canonical word orders. Head-final languages place syntactic heads after material that may generate predictions, while head-initial languages produce other storage profiles. Hsiao & Gibson (2003) reported an object-relative advantage in Chinese materials, an important challenge for a simple universal subject-relative preference. Later evidence from Vasishth, Chen, Li, & Guo (2013) found contrasting results across experiments and a Bayesian synthesis favoring a subject-relative advantage overall. That disagreement is scientifically productive: it forces resource explanations to specify constructions, control materials, statistical uncertainty, and the degree to which language-specific distributions alter predicted memory costs.
Cue-Based Retrieval and Similarity Interference
Dependency resolution frequently requires the retrieval of a previously processed phrase. When a verb imposes requirements on its subject or a gap needs an antecedent, a comprehender must identify the relevant representation despite intervening expressions. Lewis et al. (2006) articulated a model in which retrieval is content-addressable, sensitive to interference from similar items, and embedded in a sharply restricted attentional focus. A key prediction is that representational similarity can be costly even when simple temporal distance is relatively short. This shifts explanatory attention from a general capacity limit to the particular cues a linguistic operation uses to retrieve its intended argument.
Experimental retrieval interference can be studied by manipulating whether intervening nouns share relevant semantic or syntactic features with the target. Van Dyke & McElree (2006) provided evidence that additional material may create retrieval difficulty through competition rather than merely by increasing the number of words to remember. The study's interpretation depends on how distractors are constructed and whether their properties overlap with retrieval cues. Memory theories therefore gain empirical leverage from factorial manipulation of similarity and distance, with the recognition that these variables can interact. An effect of interference should not automatically be attributed to a single locus, because encoding and retrieval processes may both contribute.
Agreement attraction provides a complementary diagnostic. A sentence with a singular subject and a plural intervening noun can sometimes yield attenuated disruption at an erroneously plural verb, even though the actual grammatical controller is singular. Wagers et al. (2009) examined this phenomenon in self-paced reading and connected the pattern to representations and retrieval processes in comprehension. These findings illustrate why apparent grammatical acceptance can reflect cue-driven retrieval of a competitor rather than successful construction of a different grammatical rule. The broader scientific problem is to discriminate retrieval dynamics, probabilistic structural expectations, and strategic reinterpretation with detailed time-resolved and accuracy measures.
Filler–Gap Dependencies and Active Search
Long-distance dependencies connect expressions that are separated in the overt word sequence. In an interrogative such as 'Which document did the editor say the researcher revised?', the fronted phrase must be associated with the object position of a later verb. An incremental parser can actively seek a licensed gap rather than wait until the clause is complete. Aoshima et al. (2004) investigated how such expectations operate in Japanese, using self-paced reading and sentence completion tasks to test the association of fronted wh-expressions with embedded clauses. The results demonstrate that active dependency formation is sensitive to the grammatical organization of the target language, which may place heads and scope-marking material in positions quite different from English.
The causal interpretation of a filled-gap effect requires careful controls. Readers may experience disruption when a position anticipated as an empty dependency site is visibly occupied by another expression, but this disruption can also reflect lexical surprisal or a violation of a locally preferred clause structure. Contrasting occupied and unoccupied candidate positions, and obtaining independent probability estimates, helps distinguish these mechanisms. Active search models should specify which grammatical islands prohibit gaps and how the parser learns those constraints. Levy (2008) offers a probabilistic framework for examining structurally conditioned expectations, while Lewis et al. (2006) provides a retrieval account for linking a distant filler to a later licensed location.
Different languages provide natural experiments in the ordering of syntactic information. Japanese, Chinese, and English vary in head direction, relative-clause position, case and morphosyntactic signals, and the locations where ambiguity is resolved. Aoshima et al. (2004) makes clear that an English-centered account of filler–gap search cannot simply be transplanted without a language-specific grammar. A general psychological claim must specify which mechanism is expected to generalize: anticipatory creation of dependency representations, retrieval of a filler, integration cost, or the influence of distributional expectations. Cross-language tests are especially powerful when they distinguish these shared computational commitments from superficial differences in local word order.
Attachment Decisions, Modifier Relations, and Phrase Structure
Structural attachment is the operation of deciding where a newly encountered constituent belongs in an evolving representation. An adjective, relative clause, or prepositional phrase may modify one of several available nouns, and the parser must use grammatical constraints and information about plausible relations. Traxler et al. (1998) studied adjunct attachments with eye-tracking, manipulating whether a modifier could attach to different positions in a complex noun phrase. Their experiments demonstrate that modifier attachment is an independent empirical problem and that patterns observed in one class of ambiguity should not be assumed to transfer directly to every other class.
Attachment preference may reflect the recency of a candidate site, relative structural complexity, prosodic grouping, semantic fit, and the distribution of alternatives in the language. These variables interact in ordinary sentences, making a single-factor theory difficult to evaluate without tightly matched materials. The experimenter should distinguish whether an ambiguity is syntactic, semantic, or jointly determined by both. MacDonald et al. (1994) explains why lexical experience can bias a structure, whereas Frazier & Rayner (1982) tests preferences proposed as general parsing principles. A useful analysis compares predictions at the earliest disambiguating point and examines downstream rereading rather than relying on an end-of-sentence preference judgment.
Modifiers also reveal that attachment and interpretation can dissociate. A reader may initially prefer one noun as the attachment site, revise the syntactic representation after a disambiguating phrase, and nonetheless retain an inference associated with the initial site. Christianson et al. (2001) provides a general demonstration of persistence in earlier role assignments, although its stimuli investigate a related garden-path problem rather than every kind of modifier attachment. To test attachment specifically, experiments should ask independent questions about the grammatical host of the modifier, the intended referent, and thematic plausibility. These outcomes connect sentence processing to reference resolution without collapsing one into the other.
Thematic Roles, Plausibility, and Sentence-Level Meaning
Comprehension requires more than specifying constituent boundaries: the sentence must support a model of events and participants. Verbs constrain which arguments are plausible agents, patients, experiencers, or themes, and nouns contribute animacy and semantic features. Trueswell et al. (1994) showed that such thematic information can influence the resolution of reduced-relative ambiguities during reading. The important theoretical claim is that role information may constrain structure early enough to alter an apparent garden-path effect. Researchers must separate the availability of a semantic cue from the processing system's willingness to use it, because both vary across tasks and linguistic environments.
Temporarily assigned roles can remain active after structural reanalysis. Christianson et al. (2001) used garden-path constructions to investigate whether participants ultimately remove mistaken thematic assignments; the evidence showed that some misleading relations lingered alongside features of the correct interpretation. Ferreira et al. (2002) interprets such effects in terms of representations that are sufficient for many comprehension goals but less exact than the input's full compositional structure. The results motivate a distinction between the parser's local structural decisions, the event representation constructed from those decisions, and the answers a comprehender produces in a subsequent test. Each is a separate dependent measure and each can follow a different time course.
The relevant limits of semantic plausibility are equally important. An event may be highly plausible while violating the grammatical relations expressed by the sentence, and a grammatical sentence may describe a semantically surprising event. Comprehenders can rely on plausibility when exact structure is difficult or when the task encourages rapid understanding. Ferreira et al. (2001) documents confident misinterpretation after apparent repair, illustrating why correct grammatical classification and correct thematic understanding must be measured separately. Future work should manipulate plausibility independently of structure while controlling visual, lexical, and discourse cues; otherwise, an apparent semantic effect may be attributed to the wrong representational level.
Cross-Linguistic Variation and Relative-Clause Complexity
Relative clauses are among the best-studied tests of competing processing mechanisms because their extraction position, embedding depth, and word order can be varied systematically. In many English materials, subject relatives are easier to process than object relatives. However, this contrast can reflect integration difficulty, retrieval interference, information structure, and differing prior probabilities of constructions. Gibson (1998) offers a locality-based explanation of several relative-clause asymmetries, while Levy (2008) shows how an expectation model can generate alternative predictions. An observed asymmetry therefore defines a pattern that a theory must explain, rather than a unique signature of one memory mechanism.
The Chinese evidence is an especially important example of scientific disagreement. Hsiao & Gibson (2003) reported an object-relative advantage in a self-paced reading study and argued that differences in relative-clause and head ordering constrain universally stated complexity claims. Vasishth, Chen, Li, & Guo (2013) conducted additional experiments and a Bayesian synthesis of relevant studies, finding evidence that overall favored a subject-relative advantage while individual experiments differed. These results should not be compressed into a statement that Chinese uniformly reverses or uniformly reproduces English preferences. The conclusion depends on materials, measures, analysis and the statistical distribution of findings, and the unresolved contrasts remain informative.
Cross-language comparison must take seriously how grammatical dependencies are encoded. Languages vary in case marking, agreement, head direction, the placement of relative clauses, and whether a cue appears before or after the relevant noun phrase. Reading skill and exposure to particular constructions also vary across populations. A scientifically useful comparison identifies equivalent processing questions, then constructs stimuli that respect each language's grammar rather than imposing a word-for-word translation. The conjunction of Aoshima et al. (2004), Hsiao & Gibson (2003), and Vasishth, Chen, Li, & Guo (2013) shows why linguistic diversity can discriminate universal hypotheses about memory and expectation from local patterns arising from the surface form of English sentences.
Speech, Visual Context, and Prosodic Constraints
Spoken sentence processing operates under acoustic timing, prosodic grouping, and a listener's continuously changing environment. Speech is not a series of uniformly delimited written words; the listener must map variable acoustic evidence onto lexical candidates and syntactic relations while the input unfolds. Visual context can immediately reduce referential uncertainty. In experiments reported by Tanenhaus et al. (1995), listeners' eye movements during spoken instructions showed that information in the scene can influence online language comprehension. The theoretical implication is that syntactic decisions are sometimes made with access to task-relevant, nonlinguistic constraints before the utterance is complete.
Prosodic grouping also supplies information about clause boundaries and attachment. Pauses, timing, pitch movements, and prominence can signal structural organization, but their values are probabilistic, task dependent, and sensitive to language and speaker. The presence of a prosodic cue does not establish that listeners always treat it as an inviolable grammatical instruction. The methodological challenge is to manipulate acoustic form while preserving lexical content and to test whether the cue alters first interpretation, repair cost, or both. Eberhard et al. (1995) reviewed evidence that lexical, sublexical, and prosodic information can interact with visual referential context in naturalistic spoken language tasks.
The modality distinction matters when evaluating claims about a single human parser. Eye movements during visual reading index fixations on orthographic material, while visual-world eye movements index attention to objects under spoken instructions. Both are time-resolved measures, but they have different causal relationships to the hidden computations of parsing. Vasishth, von der Malsburg, & Engelmann (2013) cautions against interpreting eye movements as a direct transcription of parser state. A robust theory should use comparable structural manipulations across reading and listening, account for input timing, and specify which effects arise from grammar, discourse, and prosody rather than from differences in how a dependent variable is recorded.
Self-Paced Reading and Experimental Timing
Self-paced reading provides a relatively accessible method for locating processing costs along a sentence. In the moving-window variant, the participant reveals successive words or regions, and the time between actions becomes the reading-time measure. Properly designed comparisons can isolate costs at the disambiguating region and examine downstream spillover effects. Ferreira & Clifton (1986) used this approach alongside eye-tracking, while Hsiao & Gibson (2003) and Vasishth, Chen, Li, & Guo (2013) used it to investigate relative-clause complexity. The method's strength is experimental control over the sequence of input; its weakness is that an overt advance response adds a motor and decision component to the processing interval.
Region length, punctuation, lexical frequency, presentation mode, and participant strategy all influence reading time. A word-by-word task does not capture ordinary parafoveal preview or voluntary backward eye movements, and it may change the reader's pacing. Researchers therefore need several lexical controls and a principled account of where an effect is expected. Trueswell et al. (1994) emphasized that correction for region length can be problematic when scoring regions are small. A meaningful analysis must specify whether increased latency reflects the first encounter with disambiguating evidence, spillover from a preceding phrase, or a later response criterion introduced by the experimental task.
Statistical analyses should preserve participant and item variability rather than treating every word reading as an independent observation. Accuracy checks test engagement but can also change how carefully people parse, so their frequency and wording should be documented. Model comparisons should test directional hypotheses using preregistered contrasts when feasible, with uncertainty intervals rather than only declarations of significance. Demberg & Keller (2008) illustrates a complementary naturalistic approach in which naturally occurring reading times are analyzed against formal predictors. Together, controlled and corpus-based measures clarify what self-paced reading can reveal about structure and what it leaves open about the mechanism generating the response.
Eye-Tracking and the Dynamics of Reading
Eye-tracking captures fixation sequences, regressions, and the spatial distribution of reading behavior. The method was central to Frazier & Rayner (1982), which located the cost of an unexpected syntactic continuation near its point of disambiguation and examined selective rereading. Researchers often analyze first fixation duration, gaze duration, go-past time, total reading time, and regression probability. These measures have different relations to lexical access, structural revision, and subsequent rereading. Their usefulness depends on defining the region of interest before observing the results and distinguishing planned comparisons from exploratory measures selected after seeing an effect.
Eye movements are generated by a perceptual-motor control system interacting with language processing. Difficulty in parsing can affect the eyes, but fixations do not map one-to-one onto syntactic operations. Vasishth, von der Malsburg, & Engelmann (2013) explains why oculomotor models, spatial patterns, and word-based summary measures must be considered jointly. Engelmann et al. (2013) integrated theories of syntactic difficulty with eye-movement control, underscoring the value of models that predict observable fixation histories rather than linguistic processing times in isolation. A theory that fits only a late aggregated reading-time measure can miss important differences among early commitment, delayed interference, and targeted repair.
Naturalistic eye-tracking corpora broaden the inference from hand-crafted sentences to varied genres and contexts. In Demberg & Keller (2008), the Dundee corpus included reading data from ten participants across a large body of newspaper text; the analysis compared dependency locality and surprisal predictors. Such datasets supply lexical and structural diversity while introducing correlations among word length, frequency, contextual probability, punctuation, and discourse placement. Robust conclusions require sensitivity analyses and out-of-sample prediction. Experimental and corpus approaches can then answer complementary questions: whether a causal manipulation changes behavior under controlled conditions and whether the proposed mechanism generalizes to ordinary reading.
Functional Neuroanatomy and the Limits of Localization
Neuroimaging and neurophysiology identify networks recruited during syntactic and sentence-level language tasks. A major theoretical issue is how posterior temporal and inferior frontal regions participate in structural representation, retrieval, control, and integration. Friederici (2011) presents a broad review of the human language-processing architecture, while Matchin & Hickok (2020) synthesizes evidence for the cortical organization of syntax and distinguishes computations associated with different candidate regions. These accounts contribute hypotheses about neural pathways and functions, but the precise allocation of syntax-related computations remains an active research issue because many tasks also manipulate meaning, memory, and executive demand.
Functional magnetic resonance imaging measures changes in blood oxygenation associated with neural activity across relatively broad time windows. It can demonstrate that a brain region is more responsive in one experimental condition than another, but its temporal resolution cannot straightforwardly reconstruct the sequence of phrase attachments within an individual sentence. Fedorenko et al. (2012) distinguished neighboring language-selective and domain-general responses in a region historically grouped under Broca's area. The finding is important for interpreting sentence difficulty: greater activity in a frontal region may reflect linguistic representation, domain-general task demands, or both, depending on the location and contrast.
Converging causal and correlational methods should distinguish necessity from participation. Lesion studies, stimulation, invasive recordings where ethically and clinically available, and individual-subject functional localization each offer different advantages. The strongest inference arises when a precise theoretical computation, such as dependency retrieval or structure building, predicts a selective neural effect while lexical and semantic complexity are independently controlled. Matchin & Hickok (2020) provides a framework for developing such predictions; Lewis et al. (2006) supplies a cognitive-memory account that can be tested against neural evidence. Neurobiology can constrain the architecture of parsing without implying that a single anatomical region exclusively implements all syntactic operations.
Methodological Validity, Replication, and Model Comparison
Sentence-processing experiments face recurrent problems of construct validity. Reading-time differences can result from lexical surprise, syntactic attachment, referential revision, memory competition, or response strategies, and a single manipulation may alter several of these simultaneously. Ferreira & Clifton (1986) and Trueswell et al. (1994) offer an instructive historical comparison because they make similar theoretical questions sensitive to different controls. Modern designs should independently norm plausibility, structural expectations, lexical probabilities, and question difficulty. Reports should preserve all planned measures and explain why a particular region or time window is theoretically diagnostic.
Replicability also depends on how linguistic items and participants are sampled. A strong claim about the human parser should survive more than one small collection of sentences and should be examined across groups with different reading experience and language background. Vasishth, Chen, Li, & Guo (2013) illustrates how additional experiments and an evidence synthesis can produce a more qualified account than a single striking cross-language finding. Nieuwland et al. (2018) demonstrates the value of coordinated multi-laboratory work in testing claims about probabilistic prediction. Neither example provides a universal numerical correction for all parsing research; each reinforces the need to report uncertainty and the boundaries of the sampled population.
Formal comparison is preferable to post hoc compatibility. A serious parser model should generate predictions over words, items, participants, or neural responses using parameters fitted on distinct data. Demberg & Keller (2008) compared locality and surprisal on corpus reading measures, while Engelmann et al. (2013) explored the interaction between syntactic processing and eye control. Models may predict the same aggregate cost for different internal reasons, so diagnostic designs should choose stimuli where their trajectories diverge. Evaluation should include predictive fit, parameter stability, sensitivity to corpus choice, and whether the model reproduces important failures of comprehension as well as successful reading.
New Evidence on Surprisal and Syntactic Disambiguation
Recent research makes the distinction between broad word-level predictability and structural disambiguation especially consequential. Huang et al. (2024) assembled a large benchmark of reading times for syntactically complex English sentences and evaluated surprisal estimates derived from neural language models. According to the reported results, the models substantially underestimated processing difficulty for many constructions, failed to recover relative difficulty among several garden-path types, and explained little of the item-by-item variation of interest. The finding challenges a strong sufficiency claim: accurate estimates of upcoming-word probability do not automatically explain every cost produced by abandoning a particular parse. The result should be interpreted at the level of the tested constructions, models, and measurements.
A different evidentiary question is whether lexical surprisal predicts some reading-time variation across languages. Wilcox et al. (2023) investigated eleven languages spanning five language families and reported cross-linguistic support for effects of surprisal, contextual entropy, and a linear relation between surprisal and processing time within their operational framework. These findings are compatible with the more targeted failures observed by Huang et al. (2024): a predictor can explain a broad distribution of word-by-word behavior while remaining inadequate at strongly diagnostic syntactic revisions. The scientifically useful distinction is between explaining variance across many ordinary words and explaining the additional cognitive demands of a disambiguating structure.
Future benchmarks should place general naturalistic text and controlled syntactic ambiguity within a common preregistered evaluation framework. Models should be scored on their ability to predict reading-time magnitude, regional distribution, contrasts between constructions, and differences between items, rather than a single global correlation. Demberg & Keller (2008) already demonstrated that naturalistic corpora distinguish at least some locality and expectation effects; Huang et al. (2024) extends this methodological principle with a dedicated disambiguation benchmark. The issue now is to identify which mechanisms generate the residual costs that persist after controlling neural language-model surprisal.
Model Scale, Experience, and the Human Reading-Time Criterion
A more powerful language model is not automatically a closer psychological model of a human reader. Oh & Schuler (2023) analyzed surprisal estimates from language-model families of differing scales and reported that larger Transformer models with lower perplexity could provide a worse fit to human reading times. Their analysis identified specific residual patterns, including discrepancies at named entities and function words. This is an important counterexample to the assumption that optimizing next-word prediction by itself steadily improves cognitive explanation. Perplexity measures the efficiency of a model on a distribution of text; human reading times reflect expectations conditioned by learning history, attention, task demands, and memory constraints.
The cross-language and individual-experience aspects of this question require separate validation. Boeve & Bogaerts (2025) systematically evaluated surprisal measures from Dutch language models for sentence, paragraph, and book reading, emphasizing the effects of language-model choice and contextual scale on fit to eye-movement data. Such analyses extend the modeling target beyond familiar English psycholinguistic materials. Yet matching a population average is different from modeling a specific reader, and improved statistical fit can result from calibration or corpus selection without demonstrating that the same internal representations are present in both systems.
A rigorous comparison should reserve distinct datasets for estimating model parameters, selecting architectures, and evaluating generalization. It should report how probabilities are tokenized, how subword pieces are combined, whether contextual windows include relevant preceding material, and whether the text used for evaluation could have appeared in model training. Oh & Schuler (2023) shows that model scale and performance are analytically distinct, while Boeve & Bogaerts (2025) supplies a recent case in which linguistic context is examined across reading levels. These results direct attention toward controlled differences in experience and representation, not toward a single model leaderboard as a sufficient account of comprehension.
Integrating Memory and Prediction in Transformer-Based Accounts
Expectation and retrieval theories have often been compared as if they were mutually exclusive explanations of sentence-processing difficulty. A recent computational proposal brings them into the same architecture. Ryu & Lewis (2025) analyzed Transformer attention in relation to cue-based retrieval and argued that attention distributions show signatures of similarity-based interference. In the study's reported modeling approach, an entropy measure derived from word-by-word attention patterns and word surprisal together predicted aspects of reading time. This result provides a candidate computational bridge between the graded expectations emphasized by Levy (2008) and the retrieval mechanisms emphasized by Lewis et al. (2006).
The theoretical gain depends on the interpretation of the computational variables. Transformer attention weights are measurable model quantities; whether they correspond to a cognitively realistic retrieval operation requires independent tests rather than an assumption based on vocabulary such as 'attention' or 'memory.' Similarity interference, retrieval success, and the time needed to integrate a syntactic dependency should be manipulated in controlled sentences. A successful account would predict both general reading-time trends and the distinct outcomes of targeted grammatical illusions, long-distance dependencies, and garden-path reanalysis. The evidence in Ryu & Lewis (2025) is an important mechanistic proposal but does not settle the question of biological implementation.
These comparisons establish a more precise role for artificial systems in psycholinguistics. Artificial models can supply quantified expectations, representations, and internal diagnostics that can be experimentally perturbed. They also reveal failures that are informative for cognitive theory: Huang et al. (2024) documents disambiguation errors of broad surprisal predictors, while Oh & Schuler (2023) shows that improved language-model perplexity does not guarantee improved human-time prediction. The appropriate outcome of cross-system research is an evaluated mapping between defined operations, observed behavior, and representational mechanisms, alongside uncertainty about which operations are shared.
Recent Research Synthesis and Future Experimental Design
The contemporary literature sharpens a central distinction: sensitivity to probability is well supported in numerous reading environments, while particular structural disambiguations can produce additional difficulty that probability-only accounts fail to predict. The cross-linguistic evidence of Wilcox et al. (2023) and the English disambiguation benchmark of Huang et al. (2024) are informative precisely because their targets differ. A future experiment can test both in the same population by combining natural reading, deliberately ambiguous constructions, and independent lexical expectation estimates. Interpretation must attend to which predictor is theoretically allowed to change at a disambiguating word, and which residuals indicate retrieval, reanalysis, or another operation.
New computational architectures make it feasible to derive separate expectations about probability, retrieval interference, and the dynamics of memory access. Ryu & Lewis (2025) supplies an example of how attentional patterns inside a Transformer can be related to the memory-based tradition. Boeve & Bogaerts (2025) reminds researchers that language and context scale are essential methodological variables, and Oh & Schuler (2023) cautions that stronger language prediction does not guarantee a better psychological fit. Across these approaches, theoretically meaningful tests should require held-out materials and multiple behavioral measures rather than parameter tuning until one reading-time dataset is reproduced.
The resulting research agenda prioritizes transparent stimulus construction, explicit parsing representations, statistically reproducible model comparisons, and generalization across languages. Controlled experiments remain essential for identifying the causal role of syntactic manipulations; large corpora help identify whether candidate predictors retain value in normal comprehension. Physiological measures add temporally sensitive constraints but do not directly reveal an entire parse. A fundamental science of sentence processing integrates these forms of evidence without treating any one measure as a direct observation of the parser. The most important goal is to explain the time course of structure building and its failures at precisely those points where reasonable theories predict different outcomes.
Agreement, Grammatical Illusions, and Retrieval Diagnostics
Agreement processing reveals how a reader can register the wrong representation during syntactic evaluation. The number or other morphosyntactic properties of a verb must be reconciled with a grammatically licensed controller, often separated from the verb by intervening nouns. In configurations where an intervening plural noun appears inside a phrase headed by a singular subject, an ungrammatical plural verb can sometimes be processed with unexpectedly little disruption. Wagers et al. (2009) examined such attraction effects and argued that they are informative about comprehension mechanisms rather than merely errors of production. The experimental contrast is especially useful when the same nouns appear in conditions that differ in agreement well-formedness.
A cue-based retrieval approach explains a grammatical illusion by proposing that the verb triggers a search for a subject representation using features such as number, structural position, and other contextual information. A competing noun that partially matches those cues can intrude, even when its syntactic role makes it an inappropriate controller. Lewis et al. (2006) provides the broader architecture of content-addressable memory, while Van Dyke & McElree (2006) identifies similarity-based interference as a factor in sentence comprehension. An alternative explanation may appeal to encoding distortion or a noisy representation of the dependency. A convincing study separates attraction during representation building from interference at the point of retrieval, using multiple grammaticality conditions and carefully matched interveners.
The phenomenon connects parsing accuracy, response criteria, and the problem of incomplete interpretations. An unexpectedly short reading time at a grammatical violation may reflect partial retrieval of a misleading competitor rather than grammatical endorsement of the entire sentence. Christianson et al. (2001) shows why the interpretation left behind by an earlier analysis should be measured separately from online processing cost. Comparing accuracy questions, well-formedness judgments, and reading times can identify dissociations that are otherwise hidden in one dependent measure. The broader implication is that successful performance on one syntactic diagnostic does not guarantee an equally precise representation on another.
Predictive Models, Artificial Language Systems, and Explanatory Limits
Computational language models provide a way to quantify expectations over upcoming words and, in some implementations, over syntactic structures. Hale (2001) and Levy (2008) formulated expectation-based accounts in which structured probability distributions generate processing predictions. A parser can therefore be evaluated against human reading time using surprisal without claiming that a human sentence representation is identical to the internal state of the computational system. The methodological question is whether a model's expectations predict independently measured behavioral outcomes after controlling lexical frequency, length, and other sources of variance. Prediction should be examined across ambiguous and unambiguous constructions rather than reduced to aggregate performance on an arbitrary text corpus.
The Artificial Era makes comparison between human and artificial sentence processing a coherent empirical research program. Artificial systems can be tested for attachment preference, sensitivity to case and agreement, structural generalization, and responses to garden-path disambiguation. However, a generated correct answer is a task outcome; establishing similarity in processing mechanism requires more discriminating evidence, including controlled perturbations, generalization to unseen structures, and transparent measures of the artificial system's internal representations. MacDonald et al. (1994) and Lewis et al. (2006) illustrate that even for human readers, similar outputs may result from different procedures. Cross-system comparison therefore requires carefully matched explanatory targets rather than direct inference from a model's fluent language output to a human-like parser.
Human language-processing theories can make artificial model evaluation stronger by supplying adversarial constructions and independently motivated contrasts. Locality accounts Gibson (1998), cue-based retrieval accounts Lewis et al. (2006), and expectation accounts Levy (2008) predict partly different behavior under changes in distance, interference, and probability. A machine's responses to these manipulations can establish specific functional regularities, while separate interventions and model-internal analyses test candidate mechanisms. The resulting research extends cognitive science into comparative architectures of language processing, with clear operational distinctions between observable performance, algorithmic explanation, and claims about a system's broader psychological organization.
Variation Across Tasks, Readers, and Contexts
Reading and listening strategies vary with the reader's goals, the linguistic material, and the reliability of available information. A participant who expects difficult grammaticality judgments may allocate resources differently from someone reading a natural narrative for meaning. Ferreira et al. (2002) argues that comprehension can settle for representations sufficient for current purposes, while Christianson et al. (2001) provides direct evidence of residual garden-path interpretations under comprehension testing. The relevant research task is to determine which aspects of an analysis remain stable across instructions, and which change when comprehension questions, response deadlines, or error consequences alter the value of detailed reanalysis.
Experience with a language affects the relative probability of syntactic continuations, the interpretation of verb subcategorization cues, and the ease of using morphosyntactic markers. Roland et al. (2007) documents distributional differences among grammatical structures in language data, offering a foundation for studying learned expectations. However, a corpus distribution is not identical to any individual's learning history, and a statistically typical speaker is an abstraction over heterogeneous exposure. Controlled experiments need to measure or approximate the relevant language experience when evaluating probability-based explanations. Cross-language evidence Hsiao & Gibson (2003) Vasishth, Chen, Li, & Guo (2013) further demonstrates that conclusions drawn from one grammar and one set of reading materials need explicit tests of generalization.
Individual variation should be characterized through reliable measures rather than post hoc grouping into convenient categories. Working-memory performance, proficiency, reading fluency, and response strategy can all plausibly influence outcomes, but correlations across such measures require replication and a clearly specified causal interpretation. The mechanisms proposed in Lewis et al. (2006) suggest concrete experimental interventions in memory interference, while Demberg & Keller (2008) indicates how corpus-based models can test predictions over many naturally occurring word positions. Future studies will benefit from hierarchical models that separate differences among readers, items, and constructions while preserving the uncertainty surrounding individual-level estimates.
Integrative Scientific Discussion and Unresolved Problems
The main theoretical frameworks explain overlapping aspects of sentence processing. Serial structural strategies predict local attachment choices and the need to revise them Frazier & Rayner (1982); constraint-based competition explains sensitivity to lexical and semantic support MacDonald et al. (1994) Trueswell et al. (1994); surprisal formalizes costs associated with low-probability continuations Levy (2008); locality and retrieval accounts emphasize integration demands and interference Gibson (1998) Lewis et al. (2006). Their common empirical target is the distribution of effort and interpretation across a sentence. A substantive comparison asks where their quantitative predictions diverge under controlled manipulation, rather than counting how many familiar findings each theory can redescribe after the fact.
At least four problems remain open. First, the precise timing of contextual influence varies with materials and measures: Ferreira & Clifton (1986) and Trueswell et al. (1994) remain a clear historical example. Second, the relation between structural repair and final meaning is unsettled by simple accuracy measures, as the findings of Christianson et al. (2001) demonstrate. Third, cross-language relative-clause preferences are sensitive to how an asymmetry is measured, as illustrated by Hsiao & Gibson (2003) and Vasishth, Chen, Li, & Guo (2013). Fourth, the links among reading-time, ERP, and neuroimaging effects are model dependent; Osterhout & Holcomb (1992), Frank et al. (2015), and Matchin & Hickok (2020) constrain possible explanations without mapping a unique neural signature onto every parsing operation.
A productive research strategy combines preregistered structural predictions, independent lexical and plausibility norms, controlled ambiguity manipulations, and models evaluated on unseen items. Cross-language replication and explicit reporting of participants, exclusion rules, and uncertainty support generalization. Online measures should be paired with carefully designed final-interpretation tasks because the success or failure of reanalysis is part of the subject being explained. Computational models should specify whether they predict a sequence of representations, a processing cost, or an answer, and their evaluation should match that target. These methodological priorities follow from the converging and conflicting evidence examined throughout the monograph rather than from a claim of one final universally accepted architecture.
Conclusion
Sentence processing is the real-time construction and revision of grammatical and interpretive relations from unfolding language. Syntactic parsing contributes structure: attachment, dependency formation, and the selection of candidate relations among phrases. The empirical literature establishes that readers and listeners often commit to analyses before the input is complete, encounter localized difficulty when a favored interpretation fails, and sometimes carry residual meaning from an abandoned analysis into the final interpretation. Eye-movement garden-path experiments Frazier & Rayner (1982) and studies of persistent misinterpretation Christianson et al. (2001) together reveal why online structural decisions and completed comprehension must be treated as separate observational targets.
Theoretical synthesis requires several complementary mechanisms. Constraint sensitivity MacDonald et al. (1994), expectation and surprisal Levy (2008), integration cost Gibson (1998), and cue-based retrieval Lewis et al. (2006) each address different components of observed processing effort. Variation across sentences, languages, and tasks is central evidence for distinguishing their scope. The contrasting Chinese relative-clause findings discussed by Hsiao & Gibson (2003) and Vasishth, Chen, Li, & Guo (2013) demonstrate that a seemingly simple preference cannot be generalized without attention to the grammar, materials, and statistical uncertainty. Neural measures Kutas & Hillyard (1980) Osterhout & Holcomb (1992) supply additional constraints on the timing and functional organization of comprehension.
The scientific task ahead is to link formal representations to quantitatively testable processing operations, online behavioral and physiological measures, and the quality of final interpretations. This requires predictions that survive new linguistic materials, independent laboratories, and language diversity. Research on artificial language systems can contribute controlled comparative tests when performance, mechanism, and representation are investigated separately. Sentence Processing and Syntactic Parsing is therefore best understood as a coherent field of experimental, computational, and neurocognitive inquiry into how structured meaning is assembled and repaired over time.
References
- Aoshima, S., Phillips, C., & Weinberg, A. (2004). Processing filler-gap dependencies in a head-final language. Journal of Memory and Language, 51(1), 23–54. https://doi.org/10.1016/j.jml.2004.03.001
- Boeve, S., & Bogaerts, L. (2025). A systematic evaluation of Dutch large language models' surprisal estimates in sentence, paragraph and book reading. Behavior Research Methods, 57(9), Article 266. https://doi.org/10.3758/s13428-025-02774-4
- Christianson, K., Hollingworth, A., Halliwell, J. F., & Ferreira, F. (2001). Thematic roles assigned along the garden path linger. Cognitive Psychology, 42(4), 368–407. https://doi.org/10.1006/cogp.2001.0752
- Demberg, V., & Keller, F. (2008). Data from eye-tracking corpora as evidence for theories of syntactic processing complexity. Cognition, 109(2), 193–210. https://doi.org/10.1016/j.cognition.2008.07.008
- Eberhard, K. M., Spivey-Knowlton, M. J., Sedivy, J. C., & Tanenhaus, M. K. (1995). Eye movements as a window into real-time spoken language comprehension in natural contexts. Journal of Psycholinguistic Research, 24(6), 409–436. https://doi.org/10.1007/BF02143160
- Engelmann, F., Vasishth, S., Engbert, R., & Kliegl, R. (2013). A framework for modeling the interaction of syntactic processing and eye movement control. Topics in Cognitive Science, 5(3), 452–474. https://doi.org/10.1111/tops.12026
- Fedorenko, E., Duncan, J., & Kanwisher, N. (2012). Language-selective and domain-general regions lie side by side within Broca's area. Current Biology, 22(21), 2059–2062. https://doi.org/10.1016/j.cub.2012.09.011
- Ferreira, F., & Clifton, C., Jr. (1986). The independence of syntactic processing. Journal of Memory and Language, 25(3), 348–368. https://doi.org/10.1016/0749-596X(86)90006-9
- Ferreira, F., Christianson, K., & Hollingworth, A. (2001). Misinterpretations of garden-path sentences: Implications for models of sentence processing and reanalysis. Journal of Psycholinguistic Research, 30(1), 3–20. https://doi.org/10.1023/A:1005290706460
- Ferreira, F., Bailey, K. G. D., & 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
- Frank, S. L., Otten, L. J., Galli, G., & Vigliocco, G. (2015). The ERP response to the amount of information conveyed by words in sentences. Brain and Language, 140, 1–11. https://doi.org/10.1016/j.bandl.2014.10.006
- 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
- Friederici, A. D. (2011). The brain basis of language processing: From structure to function. Physiological Reviews, 91(4), 1357–1392. https://doi.org/10.1152/physrev.00006.2011
- Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68(1), 1–76. https://doi.org/10.1016/S0010-0277(98)00034-1
- 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 (NAACL 2001). Association for Computational Linguistics. https://doi.org/10.3115/1073336.1073357
- Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90(1), 3–27. https://doi.org/10.1016/S0010-0277(03)00124-0
- Huang, K.-J., Arehalli, S., Kugemoto, M., Muxica, C., Prasad, G., Dillon, B., & Linzen, T. (2024). Large-scale benchmark yields no evidence that language model surprisal explains syntactic disambiguation difficulty. Journal of Memory and Language, 137, Article 104510. https://doi.org/10.1016/j.jml.2024.104510
- Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials reflect semantic incongruity. Science, 207(4427), 203–205. https://doi.org/10.1126/science.7350657
- Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126–1177. https://doi.org/10.1016/j.cognition.2007.05.006
- Lewis, R. L., Vasishth, S., & Van Dyke, J. A. (2006). Computational principles of working memory in sentence comprehension. Trends in Cognitive Sciences, 10(10), 447–454. https://doi.org/10.1016/j.tics.2006.08.007
- 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
- Matchin, W., & Hickok, G. (2020). The cortical organization of syntax. Cerebral Cortex, 30(3), 1481–1498. https://doi.org/10.1093/cercor/bhz180
- Nieuwland, M. S., Politzer-Ahles, S., Heyselaar, E., Segaert, K., Darley, E., Kazanina, N., Von Grebmer Zu Wolfsthurn, S., Bartolozzi, F., Kogan, V., Ito, A., Mézière, D., Barr, D. J., Rousselet, G. A., Ferguson, H. J., Busch-Moreno, S., Fu, X., Tuomainen, J., Kulakova, E., Husband, E. M., ... Huettig, F. (2018). Large-scale replication study reveals a limit on probabilistic prediction in language comprehension. eLife, 7, Article e33468. https://doi.org/10.7554/eLife.33468
- Oh, B.-D., & Schuler, W. (2023). Why does surprisal from larger transformer-based language models provide a poorer fit to human reading times? Transactions of the Association for Computational Linguistics, 11, 336–350. https://doi.org/10.1162/tacl_a_00548
- 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
- Roland, D., Dick, F., & Elman, J. L. (2007). Frequency of basic English grammatical structures: A corpus analysis. Journal of Memory and Language, 57(3), 348–379. https://doi.org/10.1016/j.jml.2007.03.002
- Ryu, S. H., & Lewis, R. L. (2025). Memory for prediction: A Transformer-based theory of sentence processing. Journal of Memory and Language, 145, Article 104670. https://doi.org/10.1016/j.jml.2025.104670
- Staub, A. (2015). The effect of lexical predictability on eye movements in reading: Critical review and theoretical interpretation. Language and Linguistics Compass, 9(8), 311–327. https://doi.org/10.1111/lnc3.12151
- 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
- Traxler, M. J., Pickering, M. J., & Clifton, C., Jr. (1998). Adjunct attachment is not a form of lexical ambiguity resolution. Journal of Memory and Language, 39(4), 558–592. https://doi.org/10.1006/jmla.1998.2600
- 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
- Van Dyke, J. A., & McElree, B. (2006). Retrieval interference in sentence comprehension. Journal of Memory and Language, 55(2), 157–166. https://doi.org/10.1016/j.jml.2006.03.007
- Vasishth, S., Chen, Z., Li, Q., & Guo, G. (2013). Processing Chinese relative clauses: Evidence for the subject-relative advantage. PLOS ONE, 8(10), e77006. https://doi.org/10.1371/journal.pone.0077006
- Vasishth, S., von der Malsburg, T., & Engelmann, F. (2013). What eye movements can tell us about sentence comprehension. WIREs Cognitive Science, 4(2), 125–134. https://doi.org/10.1002/wcs.1209
- Wagers, M. W., Lau, E. F., & Phillips, C. (2009). Agreement attraction in comprehension: Representations and processes. Journal of Memory and Language, 61(2), 206–237. https://doi.org/10.1016/j.jml.2009.04.002
- Wilcox, E. G., Pimentel, T., Meister, C., Cotterell, R., & Levy, R. P. (2023). Testing the predictions of surprisal theory in 11 languages. Transactions of the Association for Computational Linguistics, 11, 1451–1470. https://doi.org/10.1162/tacl_a_00612