Chapter 4: Sentence Processing

Key Focus: Structural Ambiguity, Parsing Models, & Constraint-Based Processing

📖 Required Reading

• Chapter 4 from Introduction to Psycholinguistics: Understanding Language Science (2nd ed.) by Matthew Traxler.
  Link to Chapter 4

📖 Classroom Notes

Link to Noyrd

Overview

This chapter examines how we convert linear word sequences into hierarchical sentence structures, focusing on how we resolve structural ambiguities in real time. We compare competing models of sentence parsing - from syntax-first garden path theory to interactive constraint-based approaches - and explore how multiple cues (syntax, semantics, context, frequency) guide our interpretation of sentences. The chapter explains why certain sentences lead to processing difficulties and how we recover from misinterpretations.

Learning Goals

After studying this chapter, you should be able to:

• Explain structural ambiguity and how phrase structure trees represent different interpretations
• Analyze garden path sentences and why they cause processing difficulties
• Compare two-stage (garden path) and constraint-based models of sentence parsing
• Apply parsing heuristics (late closure, minimal attachment) to analyze sentence processing
• Evaluate the evidence for different constraints in sentence comprehension (context, verb frequency, animacy, prosody)
• Understand how visual context and referential needs influence parsing decisions
• Explain the tuning hypothesis and its cross-linguistic implications
• Describe predictive processing in sentence comprehension
• Apply sentence processing principles to explain real-world comprehension phenomena

1. Introduction to Sentence Processing

1.1 Core Goal

Convert linear words into hierarchical, meaningful structures to recover the speaker’s message. It relies on two fields:

• Syntax: Rules/cues for word relationships (word order, prepositions).
• Syntactic Parsing: How we use these cues in real time to build mental structures.

1.2 Structural Ambiguity: Why Organization Matters

Many word sequences can be grouped into phrases in multiple ways (ambiguity), leading to different meanings.

Foundational Example: “Dr. Phil discussed sex with Rush Limbaugh”

• Phrase Structure Trees: Visualize hierarchy—Interpretation 1 attaches “with Rush Limbaugh” to the verb phrase (VP); Interpretation 2 attaches it to the noun phrase (NP).

1.3 Key Evidence: Incremental Processing & Garden Path Sentences

• Immediacy Principle: We interpret words as soon as we see/hear them—no waiting for the sentence end.
• Garden Path Sentences: Ambiguous sentences where initial structure choices are later wrong (need revision).

Classic Example: “While Susan was dressing the baby played on the floor”

1. Initial Ambiguity: “dressing the baby” is parsed as “Susan dressed the baby” (attach “the baby” as object of “dressing”).
2. Disambiguation: “played” needs a subject—but “the baby” is already an object.
3. Reanalysis Cost: Reading time at “played” is 200-300 ms slower than the unambiguous control (“While Susan was dressing herself…")—extra effort to undo the initial parse.

2. Two-Stage Parsing Models (Garden Path Theory)

Lyn Frazier’s model (1979, 1987) proposes sequential stages: syntax first, then semantics/context.

2.1 Stage 1: Structural Analysis (Syntax-Only)

• Input: Only word categories (noun, verb, preposition)—no semantics/context.
• Process: Use fast, context-free heuristics to build one structure (prioritize simplicity).
• Output: Single phrase structure tree with word roles (subject, object).

2.2 Stage 2: Semantic Interpretation & Revision

• Input: Stage 1 structure + word meanings + context + world knowledge.
• Process: Assign thematic roles (agent, patient) and check coherence. If conflicting (e.g., garden path), revise the Stage 1 structure.

2.3 Core Heuristics (Stage 1)

Heuristics minimize processing effort and are assumed universal:

2.4 Limitations

1. Ignores early context/semantics: Can’t explain why story context resolves ambiguity before disambiguating words.
2. Fails cross-linguistic differences: Heuristics (e.g., late closure) aren’t universal—preferences depend on language frequency.

3. Constraint-Based Parsing Models

One-stage, parallel models (MacDonald et al., 1994; Tanenhaus et al., 1995) where all cues (syntax, semantics, context) interact simultaneously.

3.1 Core Principles

1. Parallel Activation: Multiple structures are activated at once (not just one).
2. Constraint Satisfaction: Structures compete—more cue support = higher activation.
3. No Syntax Bias: All cues influence activation from the start.

3.2 Key Constraints (Cues) & Evidence

Cues are tuned to past experience (tuning hypothesis):

3.3 Tuning Hypothesis

Parser weights cues based on past experience (e.g., “animate nouns = agents 80% of the time” → weight animacy heavily).

4. Interim Summary: Model Comparison

4.1 Predictive Processing (Extension)

Parser actively predicts upcoming words/structures (not just reacts):

• Example: “The boy will eat…” → participants look at cake (food) before hearing “cake” (uses verb semantics to predict).
• Predicts at multiple levels: phonological (sound), syntactic (category), semantic (meaning).

Quick Review Questions

1. What is structural ambiguity? Give an example.
2. Why do garden path sentences cause slowdowns?
3. What is the key difference between two-stage and constraint-based models?
4. Name two constraints that guide constraint-based parsing.
5. How does the tuning hypothesis explain cross-linguistic parsing differences?