There have been numerous models proposed to model human sentence parsing performance (Frazier & Fodor 1978, Marcus 1980, Church 1982, Abney 1989, Gibson 1991, Jurafsky 1993, Stevenson 1994). A brief discussion of these models and their limitations will be presented. This talk will introduce a model which accounts for parsing preferences, the ease of processing sentences with certain types of local ambiguity, as well as accounts for the gradedness and multi-dimensionality in the difficulty of processing center-embedded, high attachment, and strong and weak garden path sentences. Examples of sentences which the processor can account for are shown below:

  (1) a. Mary watched the Olympic trials on TV.
      b. Mary watched the Olympic trials on TV on his new color TV.
      c. Tom believed Bill thought Mary took out the cat on Monday.
  (2) a. John believes Bill.
      b. John believes Bill died.
      c. Susan put the book on the shelf into her backpack.
  (3) a. The mouse the cat the dog barked at chased is mine.
      b. men women children dogs bark at adore love are rare.
  (4) a. John gave the claim that Bill thought Mary died no credibility.
      b. I gave the girl whom you thought Bill liked a book.
      c. I gave the boy who you wanted to give the books to three books.
  (5) a. Without her donations to the charity that won failed to appear.
      b. Without her donations failed to appear.
      b. After the man drank the water proved to be poisoned.
  (6) a. The man gave the girl a ring impressed a watch.
      b. The horse raced past the barn fell.

The proposed model has a memory architecture which consists of a workspace of 3D-tree fragments for which the accessibility and immediate accessibility of its nodes are limited. An accessible node requires a unit of short-term memory, and the total number of short-term memory units is bounded. The parsing algorithm uses an ordered non-deterministic depth-first search algorithm, constructing tree-fragments using a modified head-corner parsing algorithm. Because of the unique memory architecture, the model requires, for a whole parse, only constant time per transition.

The proposed human parsing model was implemented as a computer program and will be demonstrated at this talk, illustrating how this model processes the above types of sentences.