Syntactic Reanalysis and Garden-Path Phenomena

Core Idea

In garden-path sentences (e.g., 'The horse raced past the barn fell'), initial parsing commits to an incorrect syntactic structure ('the horse raced past the barn' as the main clause) only to require reanalysis when new evidence mounts ('fell' indicates the horse is the subject of a reduced relative clause). Garden-path effects reveal that parsing is incremental and initially commits to the most frequent or simple interpretations, later correcting when evidence mandates. The revision process shows parsing is not purely bottom-up or top-down but interactive.

How It's Best Learned

Use self-paced reading where participants advance through garden-path sentences word-by-word, showing increased reading times at reanalysis points. Contrast with non-ambiguous control sentences and sentences biasing the correct initial parse to isolate reanalysis costs.

Common Misconceptions

• Assuming the parser considers all syntactically possible interpretations simultaneously; the garden-path effect shows the parser commits early.
• Treating reanalysis failures as processing deficits; garden-path phenomena occur in skilled readers and reflect optimal strategies given frequency distributions.

Explainer

From sentence comprehension and parsing, you know that the language comprehension system assigns syntactic structure to incoming words incrementally — it doesn't wait for the full sentence before beginning to build a parse. This incremental commitment is normally efficient because most sentences are unambiguous, and starting to build structure early means that by the time the sentence ends, processing is nearly complete. Garden-path sentences are the diagnostic case that reveals the cost of this strategy: they are constructed so that the most probable early parse turns out to be wrong, forcing the system to backtrack and rebuild.

The canonical example deserves careful unpacking word by word. *"The horse raced past the barn fell."* The parser begins: "The horse" → noun phrase, subject of the sentence. "raced" → simple past verb; "The horse raced..." looks like the main clause structure is underway. "past the barn" → prepositional phrase modifying the path of racing. By "barn," the parser has constructed a complete sentence: *The horse raced past the barn.* Then "fell" arrives. This is catastrophic: "fell" is a verb that needs a subject, and "the horse" is already taken as the agent of "raced." The parser must radically revise its analysis: "raced past the barn" is not the main predicate — it is a reduced relative clause (equivalent to "that was raced past the barn"), and "the horse" is not the agent of racing but the subject of the main verb "fell." The horse fell. It was the horse that was raced. This reanalysis is cognitively expensive and measurable.

Why does the parser commit to the wrong analysis in the first place? Because it is implementing a frequency-based greedy strategy. Simple active sentences ("The horse raced past the barn [and then stopped]") are far more common in natural language than reduced relative clause constructions ("The horse [that was] raced past the barn [by the jockey] fell"). The parser has internalized these frequency statistics and bets on the most probable parse at each decision point. This strategy is correct the overwhelming majority of the time — the cost of occasionally being wrong on garden-path sentences is outweighed by the efficiency gained on the far more common simple sentences. The garden-path effect is not a bug; it is the price of a well-calibrated probabilistic strategy operating on the rare case where its bet turns out to be wrong.

The reanalysis process itself is revealing. Self-paced reading studies show elevated reading times specifically at the disambiguating word ("fell") — sometimes 200–400 ms longer than on matched non-ambiguous controls. This is the measurable signature of the parser backtracking, discarding its previous structure, and rebuilding. Not all garden paths are equally costly: reanalysis is easier when the revision is local (affecting only the most recent phrase) and harder when it requires restructuring the entire sentence. Some garden-path sentences are so difficult that they remain partially unintelligible even after extended processing — readers can tell something is wrong without fully recovering the correct structure. This asymmetry tells us that reanalysis is not free, unlimited re-parsing but a costly, sometimes incomplete process.

Parsing is not purely syntactic — contextual and semantic information modulates garden-path strength. If preceding context makes the reduced relative reading more probable (you've been reading a story about horse racing), the garden-path effect is attenuated. If lexical information makes the verb more likely to appear in relative clauses (verbs with passive-favorable semantics), the effect is smaller. This interactivity shows that the parser is not an isolated syntactic module but a probabilistic system integrating multiple information sources simultaneously. The garden-path paradigm contributed to a decades-long debate about whether parsing is initially modular (pure syntax first, then semantics) or interactive (all information in parallel from the start). Current consensus favors interactive accounts: the parser is a flexible probabilistic machine, and the "commitment" to an early parse is not a hard syntactic rule but a weighted bet that stronger contextual evidence can override.