Plot Units: A Narrative Summarization Strategy
Wendy G. Lehnert Yale University
When a person reads a narrative story, an internal representation for that story is constructed in memory. One way that we can examine the contents of this memory representation is by asking the reader simple questions about the story. Typical question-answering behavior will reveal evidence for numerous inferences, causal chain constructions, and the predictive integration of information into instantiated knowledge structures ( Dyer & Lehnert, 1980; Graesser, Robertson, & Anderson, 1981; Lehnert, 1978). The complexity of an internal memory representation is reflected in part by the number of propositions that are present in memory, but not explicitly present in the original source text. For narrative texts, the ratio of inferred propositions to explicitly stated propositions is estimated to be about 8:1 ( Graesser, 1981). Although some of these inferences may be reconstructed only as needed for answering a question, we are still faced with the standard search space problems that characterize most problem areas in artificial intelligence. How can large memory representations be effectively accessed and searched by efficient retrieval heuristics? What are the organizational structures of memory that underlie internal representations for text? And finally, what sorts of human memory phenomena can be examined in an effort to shed light on issues of memory organization?
Even though question answering provides us with a method for examining the contents of a memory representation, the task of question answering does not readily yield a more global picture of the memory representation as a whole. We can only guess at how the various pieces fit together within a single structure. If we are interested in the structure of narrative memory representations, the sum-