Conceptual Analysis of Garden-Path Sentences

Michael J. PazzaniACL 1984

TL;DR

This paper introduces LAZY, a novel parser that models how humans understand "garden-path" sentences – sentences that initially seem grammatically incorrect but are actually valid. By integrating syntactic and semantic processing, LAZY can deterministically parse these sentences without needing to backtrack, mimicking human intuition and offering a more flexible approach to natural language understanding.

What Problem Does This Paper Address?

Natural language is full of sentences that can initially mislead even fluent speakers. These "garden-path" sentences, like "The horse raced past the barn fell," trick us into assigning an incorrect grammatical structure, forcing a conscious reanalysis. Traditional parsers often struggle with these, either rejecting them as ungrammatical or requiring complex backtracking mechanisms. The challenge is to build a system that can understand these sentences as humans do, recognizing the correct structure without getting "stuck" on the initial misinterpretation, especially for sentences that are semantically disambiguatable.

Methods

The core innovation is the LAZY parser, which extends conceptual analysis by explicitly representing syntactic information and enabling flexible interaction between semantic and syntactic knowledge.

  • Integrated Processing: LAZY doesn't strictly separate syntax and semantics. Instead, it uses a system of "requests" where each word has associated tests that check for both syntactic position and semantic compatibility (selectional restrictions).
  • Flexible Request Activation: Unlike some parsers that rigidly follow rules, LAZY can delay or prioritize requests based on the context. This is crucial for handling ambiguities.
  • Semantic Disambiguation: For sentences where semantics can resolve ambiguity (semantically disambiguatable garden path sentences or SDGPs), LAZY prioritizes semantic consistency. If a potential parse violates selectional restrictions, LAZY will delay committing to that parse, waiting for more information.
  • "Careless" Initial Parsing: When semantic expectations are strong, LAZY might initially make a "mistake" (like assuming "raced" is the main verb in "The horse raced past the barn fell"), but it's designed to recover. This mirrors how humans often make quick, intuitive parses.
  • Explicit Syntactic Representation: LAZY builds explicit syntactic structures, unlike some conceptual analyzers that eschew syntax.

Datasets

Not discussed in the paper. The paper uses example sentences to illustrate its points rather than formal datasets.

Key Insights and Ideas

  • Human-like Ambiguity Resolution: The paper argues that humans don't always parse perfectly; they make initial, intuitive guesses that are often semantically driven. LAZY models this by being "careless" when semantic cues are strong.
  • Semantics as a Primary Disambiguator: For sentences like "Three percent of the courses filled with freshmen were cancelled," semantic constraints (what can be "filled") help resolve the ambiguity of "filled" being a main verb versus a participle, preventing a garden path.
  • The Power of Delayed Commitment: By not forcing immediate decisions when faced with ambiguity, especially when semantics can eventually clarify, LAZY can achieve more robust parsing.
  • Reanalysis Mimics Learning: The paper suggests that LAZY's ability to reanalyze sentences after encountering an error, or its increased carefulness after processing SDGPs, mirrors how humans become better at understanding garden path sentences with exposure.

Key Results

The paper doesn't present quantitative results in terms of accuracy metrics on a benchmark dataset. Instead, it demonstrates LAZY's capabilities through qualitative analysis of various garden-path and semantically disambiguatable garden-path sentences. The key result is the demonstration that LAZY can deterministically parse sentences that confuse other parsers, and it can do so in a way that aligns with human understanding, particularly for SDGPs. It shows that by integrating semantic expectations into syntactic parsing, a more human-like and robust understanding can be achieved.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that its approach to determining when "ed" signifies a past participle versus a past tense is conservative, and removing certain conditions might impact parsing.
  • The current implementation might fail on sentences where even careful reanalysis is difficult without additional world knowledge (e.g., "The canoe floated down the river sank" requires knowing "float" can be transitive).
  • Future Work:
    • Extend LAZY to consult episodic memory during parsing, allowing for predictions based on individual objects and richer world knowledge.
  • Augment requests with an "EPISODIC" facet to incorporate this memory consultation.
  • Investigate how semantic and episodic processing occur at early stages of understanding, potentially influencing syntactic choices.

Summary updated

BibTeX
@inproceedings{pazzani1984conceptual,
  author = {Michael J. Pazzani},
  title = {{Conceptual Analysis of Garden-Path Sentences}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}
Figure