Gernsbacher (1990) argues that language processing is founded on general cognitive processes and that memory structures are an important part of language understanding. We agree but claim that Gernsbacher has failed to see two of the ramifications of this view. First, Gernsbacher's experimental paradigm, which uses decontextualized texts referring to nonexistent entities performing unconnected actions, is inconsistent with a memory-based approach. A more appropriate experimental paradigm would focus on texts rich in memory references. Second, Gernsbacher's structure- building metaphor is also inconsistent with the notion that memory underlies understanding. When memory is taken seriously, the foundation of understanding texts becomes the recognition of references to existing memory structures, not structure building.
2. Here is the obviously true part ("obviously true" means "we agree with Gernsbacher, and don't understand how anyone other than a linguist could believe otherwise"): (1) Most of the processes that are important for natural language understanding underlie other cognitive tasks. And (2) language understanding is primarily about underlying conceptual structures.
3. Here is the obviously false part: language understanding is structure building. This metaphor appears not only in the title, but in Gernsbacher's description of two important processes in the model. The first process, "laying a foundation," builds mental structures to which the second process, "mapping," attaches substructures, like adding strips to a papier-mache figure (p. 51). (The third process, "shifting," is a variant of the first.)
4. Barring some radical changes in computer hardware, these metaphors are useless for directing the implementation of cognitive models. More seriously, structure building is exactly the wrong image of what language understanding is all about. It is not even consistent with Gernsbacher's underlying mechanisms of enhancement and suppression. Most seriously, the structure building image goes hand-in-hand with an experimental paradigm, used throughout the book, that creates situations that work against the very underlying cognitive processes that Gernsbacher is looking for.
5. Consider your understanding of this example of Gernsbacher's: "Tina beat Lisa in the state tennis match." Who is Tina? Who is Lisa? What state? What tennis match? Who is telling you this and why? You have no answers for these questions. The sentence is an experimental test probe, not an act of communication. Its content goes in one ear and out the other. It is about nonexistent entities, and is not intended to connect to or affect your episodic memory.
6. Now consider your understanding of the first sentence of paragraph 1 of this review. Who are Birnbaum's colleagues? What is Language Comprehension as Structure Building the title of? Who is telling you this and why? Not only do you have answers, but those answers are important. They make the first sentence relevant to everything that follows.
7. Let's expand the sequence of events that occurred when reading the Birnbaum sentence. You were already in the process of reading this review. That process began when you recognized that (1) this was a review (reviews being things you certainly already know about), and, in particular, (2) it was a review of a book by Gernsbacher. You may or may not have had to add to your memory a person named Gernsbacher, a book named Language Comprehension as Structure Building, and an authoring relationship. When you read the first sentence, you recognized references to the book (the title), the authors of this review (our), another academic (Larry Birnbaum), a contentious statement of agreement (obviously true), another of strong disagreement (obviously false), an idiom (belabor the obvious), and so on.
8. The critical action in this process is "recognition." When you understand language (or anything), you recognize references to things you already know. You can recognize references to objects and people (Gernsbacher, our colleague), to events (he said), to argument structure (At first... On second thought...), to idioms (belabor the obvious), and so on. Even when you see something you don't know, you recognize that it's an instance of something you do know, for example, an author, a book title.
9. Interestingly, Gernsbacher's real model is already based on recognition, not structure building. The mechanisms of enhancement and suppression involve the activation of existing memory structures. In our terminology, "recognition" is the functional activity that occurs, and "activation" is one particular means of implementing recognition. Another is marker passing.
10. Our own work on AI models of natural language understanding began many years ago with structure building systems (Riesbeck, 1975, 1981). Those systems read texts and produced meaning structures. Later, story understanding systems were developed that saved those meaning structures in long-term memory (Dyer, 1983; Lebowitz, 1980) so that one story could remind the system of a previous story. Finally, we realized that having a parser building meaning structures and a separate system integrating those structures into memory was unnecessarily indirect. It was simpler, more efficient, and more plausible to have the parser access and update memory directly.
11. In our current model, called Direct Memory Access Parsing (DMAP) (Martin, 1990), phrasal patterns are attached directly to memory structures. We have memory structures not only for people (Tina, Larry Birnbaum), book parts (the title), games (tennis) and so on, but for events, such as someone beating someone at tennis (<actor> beat <opponent> at <tennis-match>), and causal sequences (When <antecedent>, <consequent>). The <...> forms refer to other memory structures, which have their own associated phrasal patterns. Descriptions of algorithms for structured recognition of concepts through their patterns can be found in Martin (1989).
12. What DMAP tries to model, and what Gernsbacher's model fails to take into account, is the role of long-term episodic memory in understanding. It is hard to remember sometimes just how much the average person knows, how specialized that knowledge is, and how quickly it can be accessed.
13. Here is a real-life example. A radio announcer began a newscast with the words "Good morning. <brief pause> Actor..." What followed? Almost everyone, once they were assured this was a real newscast, not a puzzle, said "some actor must have died or gotten very sick." Consider what this implies. These people have learned a specialized memory structure for news announcements about the deaths of celebrities. They have learned the standard phrasal templates for announcing such things, for example, "Actor <name> died today of <cause of death> in <place> at the age of <number>." They are capable of recognizing potential references to that memory structure in the span of time it takes to hear a few critical words.
14. This kind of immediate access to large bodies of generalized episodic knowledge occurs in real life all the time, because that is what communication with language is for. When a recognition-based system hears "Actor George Peppard passed away yesterday...," it quickly accesses and links its preexisting knowledge of George Peppard and of celebrities dying. Almost no structures need to be built because most of them are already there.
15. This doesn't happen in the kind of experiments that Gernsbacher and most other cognitive psychologists and psycholinguists perform, because the experimental context and the material, such as the Tina sentence, explicitly preclude references to episodic knowledge. Instead, something like structure building has to occur precisely because subjects know they are not supposed to do what they normally do, that is, modify their real memory.
16. In summary, we find much to applaud in Gernsbacher's work. We agree that language processing is founded on general cognitive processes, that memory structures are an important part of understanding, that language is for communication, and so on. However, we also believe that Gernsbacher has yet to appreciate the basic incompatibility between memory-based understanding and both the structure building metaphor and the standard test materials of psycholinguistics.
17. Here is an experiment compatible with memory-based understanding. Give subjects sentences like "A man walks into a restaurant and asks the waiter, 'Do you serve crabs here?'" Count how many of them say "I've heard that one before."
Dyer, Michael G. (1983) In-depth Understanding. MIT Press: Cambridge, MA.
Gernsbacher, Morton Ann (1990) Language Comprehension as Structure Building. Hillsdale NJ: Lawrence Erlbaum.
Gernsbacher, Morton Ann (1992) Precis of: Language Comprehension as Structure Building. PSYCOLOQUY 3(69) language-comprehension.1.gernsbacher.
Lebowitz, Michael. (1980) Generalization and memory in an integrated understanding system. Ph.D. Thesis, Yale University.
Martin, Charles E. (1989) Case-based Parsing and Micro DMAP in C.K. Riesbeck and R.C. Schank, eds., Inside Case-Based Reasoning, Lawrence Erlbaum Associates, Hillsdale, NJ.
Martin, Charles E. (1990) Direct Memory Access Parsing, Ph.D. Thesis, Yale University.
Riesbeck, Christopher K. (1975) Conceptual Analysis in R.C. Schank, ed., Conceptual Information Processing, North-Holland and Elsevier, Amsterdam.
Riesbeck, Christopher K. (1981) Micro ELI in R.C. Schank and C.K. Riesbeck, eds., Inside Computer Understanding: Five Programs Plus Miniatures, Lawrence Erlbaum Associates, Hillsdale, NJ.