Chapter 4: Sentence Processing II

Key Focus: Argument Structure, Parsing Model Limitations, & Long-Distance Dependencies

📖 Classroom Notes

Link to Noyrd

Overview

This chapter extends sentence processing by addressing three critical topics:

1. Argument Structure Hypothesis (ASH): Solves the “storage dilemma” of constraint-based models by distinguishing mandatory verb arguments (core partners) from optional adjuncts (extra details).
2. Limitations of Constraint-Based Models: Why simplicity sometimes overrides frequency, and alternative theories (Construal, Good-Enough Parsing, Race-Based) that blend two-stage and constraint-based ideas.
3. Long-Distance Dependencies: How we link separated words (e.g., the boy in “It was the boy whom the girl chased”) via gaps/traces or direct verb linking.

Learning Goals

After studying this chapter, you should be able to:

• Differentiate between arguments and adjuncts and explain their distinct processing characteristics
• Apply the Argument Structure Hypothesis to explain verb frame limitations and processing differences
• Analyze the key limitations of constraint-based parsing models and the evidence supporting them
• Compare alternative parsing theories (Construal, Good-Enough, Race-Based) and their unique claims
• Explain how we process long-distance dependencies using gaps-and-traces versus direct verb linking
• Evaluate the evidence for active filler strategy in ambiguous gap sentences
• Apply sentence processing principles to explain real-world comprehension phenomena and parsing difficulties
• Understand how different parsing theories account for both efficient processing and comprehension errors

1. Argument Structure Hypothesis (ASH)

1.1 Core Problem: The “Storage Dilemma”

Constraint-based models struggle to explain how we store endless verb structures (e.g., read can be intransitive/transitive/ditransitive). ASH solves this by splitting verb partners into arguments (mandatory, pre-stored) and adjuncts (optional, computed dynamically).

1.2 Arguments vs. Adjuncts: Key Differences

Examples

• Arguments:
  • Devour (transitive): Dr. Phil devoured the sandwich (grammatical); Dr. Phil devoured (ungrammatical—no object).
  • Put (triadic): Dr. Phil put the book on the shelf (grammatical); Dr. Phil put the book (ungrammatical—no goal).
• Adjuncts:
  • Quickly in Dr. Phil devoured quickly (optional—sentence is complete without it).
  • At the park in Dr. Phil put the book on the shelf at the park (optional—core action intact).

1.3 Verb Argument Frames (Stored in Lexicon)

Each verb’s lexical entry includes an argument frame (0–4 arguments max, cognition limits this number):

1.4 Evidence for ASH

1. Faster Processing of Arguments

Arguments are pre-stored → faster integration than adjuncts.

• Experiment (Clifton et al., 1991):
  • (51) The saleswoman tried to interest the man in the wallet (argument: in the wallet = required by interest).
  • (52) The saleswoman interested the man in his fifties (adjunct: in his fifties = describes the man).
• Result: Reading times for (51) are 100–150 ms faster—parser retrieves argument frames, computes adjuncts on the fly.

2. Argument Interpretation Bias

Ambiguous phrases are initially parsed as arguments (pre-stored) before revision.

• Example: The bully stapled a letter to Harry → to Harry is first parsed as an argument (goal for stapled) → revised if semantics conflict (odd to “staple to Harry”).
• Evidence: Longer eye fixations at to Harry—reflects revision.

3. Inference of Missing Arguments

Omitted arguments are automatically inferred (adjuncts are not).

• Experiment (Mauner et al., 1995):
  • The ship was sunk (passive, needs agent) → infer “by someone”.
  • The ship sank (intransitive, no agent) → no inference.
• Result: Purpose clauses (…to collect insurance) are 200 ms faster for was sunk—inferred agent links to the purpose.

2. Limitations of Constraint-Based Models & Alternatives

2.1 Key Limitations

1. Simplicity Overrides Frequency

Constraint-based models predict frequent structures are preferred—but simplicity (minimal attachment) wins, causing garden paths.

• Example (Pickering et al., 2000): The athlete realized her shoes somehow got left on the bus → realized usually takes sentence complements (90% frequency), but parser initially parses her shoes as a direct object (simpler) → slowdown at somehow.

2. No “Reverse” Garden Paths

Context favoring complex structures does not interfere with simple structures (contradicts constraint-based competition).

• Example (Binder et al., 2001): The criminal exiled his partner… → reading times are identical in 1-criminal (simple) and 2-criminal (complex) contexts—no interference.

2.2 Alternative Parsing Theories

1. Construal Theory (Frazier & Clifton, 1996)

• Core: Parse arguments (primary relations) sequentially (two-stage: syntax-first) and adjuncts (non-primary) in parallel (constraint-based).
• Example: Relative Clauses:
  • The daughter of the colonel who had a black dress left → who… is an adjunct (parallel processing) → semantics (dress = daughter) quickly selects correct structure → no slowdown.

2. Good-Enough Parsing (Ferreira, 2003)

• Core: Comprehenders build “good-enough” (not perfect) structures—stop parsing when meaning is clear.
• Evidence:
  • The mouse was eaten by the cheese → most paraphrase as The mouse ate the cheese (ignore passive syntax, use semantics).
  • While the hunter stalked the deer drank… → 70% answer “yes” to “Was the hunter stalking the deer?” (retain initial misparse—good enough for gist).

3. Race-Based Models (Van Gompel et al., 2005)

• Core: Multiple structures “race” to reach activation threshold—first to win is selected; reanalysis if wrong. No weighted activation (all start equal).
• Example: Garden Path Sentence:
  • Structure 1 (the baby = object of dressing) activates faster (simpler) → wins first → slowdown at played (reanalysis to Structure 2: the baby = subject of played).
• Evidence: Structural Priming: Prime sentences speed up activation of matching structures in the race.

3. Parsing Long-Distance Dependencies

3.1 Local vs. Long-Distance Dependencies

• Local: Related words are adjacent (e.g., The girl chased the boy—subject/verb, verb/object are next to each other).
• Long-Distance: Related words are separated (e.g., It was the boy whom the girl chased—the boy is separated from chased by whom the girl).

3.2 Two Theories of Long-Distance Parsing

1. Gaps-and-Traces Theory (Chomsky, 1965)

• Core: Moved words (fillers, e.g., the boy) leave a gap (placeholder) in their original position; parser links filler to gap via a “trace”.
• Parsing Steps for It was the boy whom the girl chased:
  1. Detect the boy as a filler (signaled by It was… whom).
  2. Search for a gap (after a transitive verb needing an object).
  3. Find gap after chased → link the boy to gap.
• Evidence: Cross-Modal Priming (Nicol & Swinney, 1989):
  • Filler (the boy) is reactivated at the gap → faster naming of boy or child (associate) at the gap vs. before the verb.

2. Gap-Free Theory (Pickering & Barry, 1991)

• Core: No gaps/traces—link filler directly to the verb that requires it (use verb’s argument frame).
• Parsing Steps for It was the boy whom the girl chased:
  1. Detect the boy as a potential direct object.
  2. When encountering chased (needs direct object), link the boy directly to chased.
• Evidence: Verb-Based Slowdown (Pickering & Traxler, 2001):
  • (68) That’s the pistol with which the killer shot the man (sensible: pistol = instrument for shot).
  • (69) That’s the pistol in which the killer shot the man (nonsensical: pistol ≠ location).
• Result: Slowdown at shot (not gap) for (69)—parser links filler to verb when verb is encountered, not at gap.

3.3 Active Filler Strategy (Ambiguous Gaps)

For sentences with multiple possible gaps, parser attaches filler to the first possible gap (even if wrong).

• Example (Stowe, 1986): That’s the boy that the girl liked [doubtful gap] to ignore [real gap] → parser first links the boy to gap after liked → revises at to ignore.
• Evidence: 250 ms slowdown at to ignore (revision effort).

Quick Review Questions

1. What is the difference between an argument and an adjunct? Give an example of each.
2. Why does ASH limit verbs to 4 arguments?
3. What is “good-enough parsing”? Give an example of when it occurs.
4. How do gaps-and-traces theory and gap-free theory differ in parsing long-distance dependencies?
5. What is the active filler strategy, and when does it cause processing slowdowns?