Sentence Processing II

Course Logistics and Assignment Framework

Course Structure and Schedule Overview:
- Topic: Psychology of Language, Lecture on Sentence Processing II.
- Conceptual Progression: Ambiguity constraints and difficulty factors ( $00\text{--}05$ and $05\text{--}20$ \,min), predictive language processing ( $20\text{--}40$ \,min), working memory architectures in parsing ( $40\text{--}50$ \,min), individual differences and executive control ( $50\text{--}70$ \,min), and concluding synthesis and discussion ( $70\text{--}75$ \,min).
- Associated Fine Art Motifs: René Magritte, The Blank Signature (1965), National Gallery of Art, Washington; Hilma af Klint, The Swan, No. 17 (1915); Giorgio de Chirico, The Nostalgia of the Infinite (1911), Museum of Modern Art, New York; Julie Mehretu, Stadia II (2004), Carnegie Museum of Art, Pittsburgh.
- Key Milestones:
- Assignment 2: Released and due Monday, November 2 at $4:59\,\text{PM}$ via Google Docs link on Brightspace.
- Upcoming Lectures: Thursday, October 22 covers Chapter 10.1–10.3 (Speaking I); Tuesday, October 27 covers Chapter 10.4–10.5 (Speaking II).
- Assessment: Midterm 2 takes place on Tuesday, November 10.
Assignment 2 Architecture and Evidence Requirements:
- Objective: Identify, isolate, and rigorously analyze real-world instances of linguistic ambiguity and speech errors from supplied audiovisual recordings (scripted comedy and live political addresses).
- Core Modules:
- Part 1: Lexical Ambiguity (evaluating polysemy and homophony grounded in Chapter 8).
- Part 2: Structural Ambiguity (evaluating syntactic garden-path constructions and attachment ambiguities grounded in Chapter 9).
- Part 3: Speech Errors (evaluating slips of the tongue and production mechanisms grounded in Chapter 10).
- Part 4: Meta-cognitive reflection describing research methodology and analytical workflow.
- Evidentiary Standards: Direct citation of media clip numbers accompanied by exact timestamps (e.g., Clip 2, $1:05$ ) and explicit theoretical justifications. Not all occurrences qualify; critical discrimination in selecting empirically valid tokens is an explicit grading parameter.
Exemplar Case Analysis: Headline Ambiguity Resolution:
- Target Headline: "Iraqi head seeks arms".
- Lexical and Semantic Ambiguities:
- The token "head" contains dual competitive lemmas: anatomical body part (the dominant frequency sense) versus political leader of a sovereign state (subordinate sense).
- The surrounding clausal context introduces secondary ambiguity with the token "arms", spanning anatomical upper limbs versus military ordnance/weapons.
- Contextual Coherence: The composite proposition "The leader of Iraq seeks military weapons" retains high global plausibility and aligns with standard geopolitical journalistic discourse over the literal anatomical interpretation.
- Theoretical Framework (Bottom-Up Spreading Activation Model; Sedivy, 2020, p. 600, Figure 8.5A):
- Initial Stage: Upon visual word recognition of "head", both the dominant anatomical sense and the subordinate leadership sense are automatically and concurrently activated via bottom-up spreading activation.
- Feature Activation: Activated lexical nodes transmit activation downward to corresponding semantic feature bundles (e.g., $[\text{+body part}]$ versus $[\text{+world leader}]$ ).
- Selection and Integration: Global context (e.g., register of a news publication, real-world geopolitical background knowledge, syntactic constraints, and communicative plausibility) suppresses the irrelevant anatomical node and selects the appropriate executive reading.

Predictive Mechanisms in Real-Time Sentence Processing

Incremental Processing Versus Predictive Architecture:
- Incremental parsing theories establish that human comprehenders do not defer interpretation until clause boundaries; rather, linguistic input is integrated continuously word by word in real time (demonstrated empirically by rapid shadowing latencies and immediate garden-path disruption).
- Predictive parsing asserts that the language comprehension architecture is proactive rather than purely reactive, continually generating anticipatory representations of upcoming phonological, lexical, and syntactic material.
- Sentence Completion Demonstrations:
- "The day was breezy so the boy went outside to fly ___" $\rightarrow$ Generates immediate anticipation for "a kite".
- "The pirate hides the ___" $\rightarrow$ Anticipates "treasure" based on thematic schema constraints.
- "The waitress will merrily bring the hamburger to the ___" $\rightarrow$ Anticipates "customer" or "table" driven by verb-argument structures and semantic roles.
Eye-Tracking in the Visual World Paradigm (Kamide et al., 2003):
- Experimental Methodology: Participants listened to spoken sentences while viewing an array containing target and distractor entities, including a motorcycle, a carousel, a motorcyclist in protective gear, a young girl, a mug of beer on a table, and a candy jar on a shelf.

Visual display from Kamide et al. (2003) visual world paradigm

Sentence Conditions:
- "The man will ride the motorbike."
- "The man will taste the beer."
- "The girl will ride the carousel."
- "The girl will taste the sweets."
Behavioral Findings: Prior to acoustic onset of the direct object noun (e.g., immediately upon hearing "The man will ride…" or "The girl will taste…"), participants launched anticipatory saccadic eye movements to the appropriate visual referent.
Theoretical Significance: Comprehenders do not merely compute selectional restrictions of the verb in isolation; they rapidly integrate the combined composition of the grammatical subject and the transitive verb ( $\text{Subject} + \text{Verb}$ ) to constrain expectations regarding the upcoming semantic patient.
- Electrophysiological Evidence for Fine-Grained Lexical-Phonological Prediction (DeLong et al., 2005):
Paradigm: Event-Related Potentials (ERPs) recorded during rapid serial visual presentation of sentences with high-predictability cloze continuations:
- "The day was breezy so the boy went outside to fly [a kite / an aeroplane]."
Rationale: An unpredictable noun (e.g., "aeroplane") routinely elicits an enhanced $\text{N400}$ component—a negative voltage deflection peaking approximately $400\,\text{ms}$ post-stimulus onset reflecting difficulty in semantic retrieval and integration.
Critical Finding: The $\text{N400}$ divergence occurred prior to the noun, manifesting directly on the pre-nominal indefinite article ("an" versus "a").

ERP waveforms from DeLong et al. (2005) showing N400 effects at the indefinite article and target noun

Theoretical Significance: The elevated negative amplitude to "an" demonstrates that the cognitive system does not just broadly anticipate an abstract semantic category (e.g., $[\text{+flying toy}]$ ); it pre-activates the specific upcoming phonological form ("kite"), leading to immediate violation responses when encountering the phonologically incompatible article "an".
- Surprisal Theory (Hale, 2001; Levy, 2008):
Formal Definition: A probabilistic parsing metric where the cognitive processing difficulty of an incoming linguistic element $w_i$ is directly proportional to its surprisal given the preceding context $C$ :

$\text{Surprisal}(w_i) = -\log_2 P(w_i \mid C)$

Computational Dynamics:
- Low Surprisal: Continuity with high conditional probability; minimal cognitive effort and low processing latencies.
- High Surprisal: Occurrence of low-probability continuation; induces substantial cognitive processing load and prolonged reading fixations.
Unifying Capacity: Surprisal theory unifies garden-path effects (structural reanalysis) and general lexical unexpectedness under a singular continuous probabilistic framework. Metric values systematically correlate with the biological amplitude of the $\text{N400}$ .
Determinants of Surprisal: Surprisal values are modulated not merely by raw corpus n-gram frequencies, but dynamically integrate real-world situation models and pragmatic world knowledge.
- Developmental Trajectory and Individual Differences in Prediction:
Borovsky et al. (2012): Visual world eye-tracking demonstrates that young children generate predictive saccades toward upcoming referents, but with significantly slower latency than adult cohorts.
Decoupling Age and Vocabulary: Chronological age does not drive the efficiency or speed of predictive gaze shifts. Receptive vocabulary size constitutes the critical predictive variable: larger lexicons facilitate faster real-time predictive activation in both children and adult populations.

Working Memory Constraints and Syntactic Complexity

Structural Architecture Taxing Working Memory:
- Center-embedded and object-extracted relative clauses produce disproportionate cognitive loads compared to right-branching or coordinate syntactic structures expressing identical semantic propositions.
- Comparative Clausal Formulations:
- Center-Embedded: "The mouse the cat chased keeled over."
- Canonical Alternative: "When the cat chased the mouse, the mouse keeled over."
- Severe Center-Embedding: "The administrator who the intern who the nurse supervised had accused fudged the medical reports."
- Right-Branching Alternative: "The nurse supervised the intern who accused the administrator who fudged the medical reports."
Subject-Relative Versus Object-Relative Asymmetry:
- Subject Relative (SR; computationally less demanding):
- "The senator who [gap] spotted the reporter shouted."
- Object Relative (OR; computationally demanding):
- "The senator who the reporter spotted [gap] shouted."
- Theoretical Accounts of Processing Asymmetry:
- Experience-Based / Statistical Frequency Account: Subject relatives occur with markedly higher distributional frequency in linguistic corpora; the processing mechanism aligns with probabilistic expectations, treating low-frequency object relatives as unexpected structural anomalies.
- Memory-Based Integration Distance Account (Dependency Locality Theory; Gibson, 1998):
  - Syntactic integration cost is a function of the distance between a dependent head and its structural governor, measured by the number of discourse-referential elements (new noun phrases or verbs) intervening between them.
  - In SR structures, the extraction site immediately follows the relativizer, yielding zero intervening discourse referents and low integration cost.
  - In OR structures, the subject noun phrase "the reporter" physically intervenes between the antecedent "the senator" and the transitive verb "spotted".
  - Processing consequence: The parser must maintain an unresolved long-distance dependency for "the senator" across intervening discourse referents while simultaneously opening a second dependency for "the reporter", stacking competing unresolved syntactic structures in working memory.
Similarity-Based Interference in Working Memory Retrieval:
- Principle: Working memory retrieval in language parsing is cue-based and subject to retroactive and proactive interference when stored items share overlapping syntactic, morphological, or semantic features.
- Empirical Contrast:
- High Interference: "The senator who the president summoned arrived." Both nominal elements are human, animate, singular, professional discourse referents, causing severe cue overload during retrieval at the verb "summoned".
- Low Interference: "The helicopter that the president summoned arrived." Differentiating the extracted item by animacy and semantic type ( $[\text{-animate}]$ vehicle versus $[\text{+animate}]$ human) substantially diminishes memory interference and eases clausal integration.
- Analogy: The cognitive strain of resolving high-interference object relatives mirrors reading a Russian novel in which multiple characters possess phonologically and socially confusable patronymic names.
Limits of Syntactic Center-Embedding:
- Single Clause: "The cat chased the mouse." (Cognitively trivial; baseline parsing efficiency).
- Single Center-Embedding: "The mouse the cat chased keeled over." (Challenging; introduces an embedded clause splitting the main subject from its matrix predicate).
- Double Center-Embedding: "The mouse the cat the dog bit chased keeled over." (Cognitively intractable; induces catastrophic breakdown of the parsing mechanism due to working memory depletion from holding three nested, unresolved nominal dependencies simultaneously).

Individual Differences and Executive Control in Sentence Parsing

Neural and Structural Parsing Heterogeneity:
- The presumption of an invariant, universally standardized human parsing architecture is empirically invalidated by profound individual variation in behavioral and neural processing profiles.
- Mahowald & Fedorenko (2016):
- Longitudinal functional magnetic resonance imaging ( $\text{fMRI}$ ) investigations reveal idiosyncratic, highly differentiated topography of cortical activations during sentence comprehension.
- Within-Subject Invariance: Individual activation topographies remain remarkably stable and robust across repeated testing sessions separated by substantial time intervals.
- Between-Subject Divergence: Distinct participants demonstrate substantially divergent spatial distributions of clausal processing networks across frontal and temporal regions.

fMRI activation maps across two testing sessions for individual participants from Mahowald and Fedorenko (2016)

Working Memory Capacity: The Reading Span Paradigm:
- Reading Span Test Methodology (Daneman & Carpenter, 1980): Participants visually process increasingly long sets of unrelated sentences, read each aloud for semantic comprehension, and must recall the terminal word of each sentence upon presentation of a terminal cue.
- Sentence Evaluation Activity:
- Item 1: "Despite their persistent disagreements, they managed to agree how to best educate their child." (Terminal token: child).
- Item 2: "At last, she spotted the rowboat coming across the bay, tossed about on the tall waves." (Terminal token: waves; Comprehension probe: What was spotted coming across the bay? $\rightarrow$ A rowboat).
- Item 3: "Long periods of unstructured reading and thinking sometimes lead to the most fertile ideas." (Terminal token: ideas).
- Cumulative Span Memory Output: child, waves, ideas.
- Impact on Parsing Dynamics (Just & Carpenter, 1992): Comprehension capacity splits systematically along working memory span. Low-span readers exhibit severe reading disruptions and comprehension failures on syntactically complex object-relative constructions, whereas high-span readers navigate these structural bottlenecks with negligible processing penalties.
Cognitive Control and Executive Function in Language Comprehension:
- Definition of Cognitive Control (Executive Function): The top-down neurocognitive capacity to selectively focus on goal-relevant representations while actively suppressing or inhibiting salient, irrelevant competitors.
- Linguistic Parsing as Inherent Competition: Natural language processing continuously requires the resolution of competing linguistic inputs, including homophonous lexical senses, alternate subcategorization frames, and rival syntactic attachment parses.
- The Stroop Interference Paradigm:
- Classic Test: Color-naming task where participants vocalize font ink colors while actively overriding prepotent automatic reading of incongruent orthographic color words (e.g., viewing "BLUE" printed in yellow ink, or "CHAIR", "LIBRARY", "GREEN", "SIMPLE", "RED" printed in mismatched inks).
- Syntactic Parallels to the Stroop Effect:
- Lexical Selection: "She decided not to attend the ball." The dominant lexical meaning of "ball" (a spherical sporting object) is automatically activated via bottom-up cues and must be suppressed via cognitive control in favor of the subordinate social dancing event.
- Syntactic Parsing: "The general presented copies of the report was aware of the problems." The parser automatically activates the high-frequency main-clause transitive verb interpretation of "presented", which must be overridden upon reaching the auxiliary "was" to adopt the reduced relative clause reading.
Developmental Parsing Failure: The Kindergarten Path Effect:
- Trueswell et al. (1999) Paradigm: Five-year-old pediatric subjects and adult controls were given spoken instructions to manipulate tabletop objects: "Put the frog on the napkin into the box."
- Physical Display Context: Included an empty box, a toy horse, an empty napkin, and a target frog situated atop a napkin.

Experimental setup for the kindergarten path task from Trueswell et al. (1999)

Adult Performance: Adults exhibit transient fixation shifts toward the incorrect destination (the empty napkin) upon hearing "on the napkin", but instantly recover upon hearing "into the box", correctly interpreting "on the napkin" as a modifier restricting which frog to move.
Pediatric Performance: Five-year-old children execute incorrect actions over $50\%$ of the time, typically moving the frog onto the empty napkin, and subsequently carrying it into the box.
Underlying Mechanism: Immature neurodevelopment of the prefrontal cortex in young children precludes the executive inhibitory control required to override an initial, misdirected syntactic commitment (destination argument) once committed.
- Cognitive Training and Generalization Limits:
The $N$ -Back Task Architecture: Participants observe a serial stream of characters and indicate whether the current stimulus matches the item presented exactly $N$ trials previously.

Comparison of low-conflict versus high-conflict 4-back training tasks from Hussey et al. (2017)

- Low-Conflict Condition: Displays target repeats and non-matching fillers without proactive lures.
- High-Conflict Condition: Introduces misleading lure items (e.g., an item matching an  $N-1$  or  $N+1$  position), requiring high executive inhibitory control to reject familiarity-based false alarms.

Empirical Intervention (Hussey et al., 2017):
- Participants completed $8\,\text{hours}$ of adaptive $N$ -back training spanning a period of $3\text{--}6\,\text{weeks}$ .
- High-conflict training cohorts (trained to suppress lure interference) demonstrated a significant reduction in garden-path recovery times during sentence reading.
- Low-conflict training cohorts exhibited no transfer effect on sentence parsing.
- Structural Specificity: Training-induced improvements were restricted entirely to syntactic ambiguity recovery (garden paths) and yielded zero enhancement for object-relative clause processing.
Methodological Skepticism (Simons et al., 2016): Comprehensive meta-analyses establish virtually no reproducible evidence that commercialized, non-linguistic working memory or "brain-training" programs confer meaningful, generalizable enhancements to real-world cognitive or linguistic proficiency.
- Mechanisms of Humor: Structural Garden Paths and Reanalysis:
Jokes as Garden-Path Constructs: Verbal humor frequently relies on establishing an initial mental model or syntactic parse, only to abruptly invalidate it at the punchline, requiring cognitive reanalysis and frame-shifting.
Case Studies in Humor:
- "Two fish are in a tank. One says to the other: 'How the hell do we drive this thing?'" (Shifts the lexical representation of "tank" from an aquatic glass enclosure to an armored military combat vehicle).
- "We were having so much fun jumping from the tree into the pool that we decided to add a little water." (Revises the default scenario model of a functional, fluid-filled swimming pool to a catastrophic dry basin).
- "A man put on a clean pair of socks every day of the week. By Friday he could hardly get his shoes on." (Revises the pragmatic assumption of daily sock replacement to cumulative, layered donning).
Cognitive Theory of Humor: Grounded in "expectation violation and resolution" (Hurley, Dennett, & Adams, Jr., Inside Jokes: Using Humor to Reverse-Engineer the Mind), wherein mirth and laughter serve as an intrinsic neurochemical reward for detecting and resolving cognitive framing errors.

Understanding Checks and Interactive Seminar Discussions

Comprehension Assessment 1: Cross-Linguistic Prediction:
- Context: In Spanish, "una" marks singular feminine nouns, whereas "un" marks singular masculine nouns. When reading "Caperucita Roja llevó la comida a su abuela en un…" (Little Red Riding Hood took the food to her grandmother in a…), readers strongly predict the feminine noun "una canasta" (a basket). When encountering the masculine article "un" instead, distinct neural violation responses trigger before any noun appears.
- Assessment Query: What parsing property does this immediate reaction demonstrate?
- Option (a): Readers could not fit the noun's meaning into the sentence.
- Option (b): "Un" is inherently harder to read than "una".
- Option (c): The brain is reacting to a formal grammar mistake.
- Option (d): Readers had predicted a specific noun, including its grammatical gender.
- Correct Classification: Option (d). The brain proactively anticipates specific downstream lexical candidates down to their formal grammatical inflectional properties (e.g., gender agreement).
Comprehension Assessment 2: Memory Load Mechanisms:
- Context: Contrast "The lawyer who the witness insulted left the room" (Object Relative) with "The lawyer who insulted the witness left the room" (Subject Relative).
- Assessment Query: Under Gibson's (1998) Dependency Locality Theory, what exact mechanism explains why the object-relative sentence is more difficult to process?
- Option (a): Readers must hold "the lawyer" in mind while processing another noun ("the witness") until reaching "insulted".
- Option (b): It constitutes a garden path where readers build the wrong syntactic tree and must restart.
- Option (c): "Insulted" is structurally ambiguous between a matrix verb and a reduced relative.
- Option (d): "Lawyer" and "witness" generate similarity-based semantic interference.
- Correct Classification: Option (a). Storing an open nominal dependency across intervening discourse referents introduces long integration distances and consumes working memory resources.
Comprehension Assessment 3: Cognitive Training Transfer:
- Context: A digital cognitive training program trains users to suppress distracting interference lures, mimicking the high-conflict $N$ -back protocol in Hussey et al. (2017).
- Assessment Query: If the program's outcomes replicate Hussey et al., what specific linguistic enhancement should users display?
- Option (a): Reading object relatives like "The senator who the reporter spotted shouted".
- Option (b): Recovering from garden path sentences.
- Option (c): Broad-spectrum reading comprehension gains across all sentence types.
- Option (d): Zero linguistic effects, as cognitive training lacks measurable transfer.
- Correct Classification: Option (b). High-conflict executive control training selectively facilitates the suppression of an initial, erroneous syntactic parse during garden-path reanalysis, but does not compress linear integration distances in object relatives.
Interactive Discussion Questions:
- Linguistic Typology and Garden Paths: Designing a cross-linguistic syntactic garden path requires exploiting features unique to English (such as unmarked relative clauses or identical surface forms for past tense and past participles). In morphologically rich languages featuring explicit case marking (e.g., German nominative/accusative distinctions) or rich verbal agreement, surface ambiguities are often precluded because inflectional morphology resolves the syntactic role of noun phrases instantly.
- Universal Plain-Language Writing Guidelines: To minimize parsing difficulty for broad, cognitively heterogeneous audiences, writers must systematically eliminate center-embedded clauses, shorten dependency integration distances, avoid reduced relative constructions lacking overt complementizers (e.g., utilizing "The report that was presented" instead of "The report presented"), minimize similarity-based interference between adjacent animate nouns, and avoid highly unexpected low-probability lexical continuations that produce high surprisal.
- Digital Interventions for Reading Comprehension: A theoretically sound reading-enhancement application should integrate dual targeted modules: (1) Domain-specific vocabulary enrichment to accelerate lexical access and expand the capacity for predictive parsing, and (2) Contextual reanalysis exercises that train readers to rapidly detect and recover from structural misinterpretations, rather than relying on abstract, non-verbal memory games that fail to transfer to language processing.