Garden-Path Sentence Parser

Garden-path sentence

Playback

Word

–

Reading time

–

Cumulative time

0 ms

Parser state

Idle

How it works

A self-paced reading experiment reveals a sentence one word at a time and records how long a reader dwells on each word before advancing — dwell time is a proxy for processing difficulty. This simulator plays that experiment on classic garden-path sentences: at every word the incremental parser attaches it to the cheapest, most likely structure it can (minimal attachment / late closure), so it commits early to an analysis that later turns out wrong.

When the next word is incompatible with that committed structure — "fell" after "the horse raced past the barn", "the boats" after "the old man" — the parser must throw away part of its tree and rebuild it from scratch. That reanalysis is expensive: reading time spikes sharply at the disambiguating word (and often lingers into the next one, a "spillover" effect seen in real self-paced reading and eye-tracking data), then returns to baseline once the correct structure is in place.

Word stream — words appear one at a time at the simulated reading pace; the current word is highlighted, the disambiguating word flashes red.
Reading-time bars — a live histogram of simulated per-word dwell time; the spike marks the reanalysis cost.
Parse tree — builds the parser's first-choice (garden-path) structure, then visibly re-roots and re-labels itself into the correct structure once the crash word forces reanalysis.

Real-world relevance: this incremental-commitment-then-repair pattern is central to psycholinguistic models of human sentence processing (Frazier's garden-path model, surprisal-based accounts) and explains measurable slowdowns in reading experiments, eye-tracking studies and ERP N400/P600 responses.

Idle

Attached node Garden-path structure Correct (reanalyzed)