Defining and Demonstrating Capabilities for Experience-Based Narrative Memory

摘要

Researchers at MIT have constructed a novel proof-of-concept demonstration of memory-driven narrative structuring of information. Their demo system is a confederation of three systems they are developing: The Story Workbench semi-automatic annotation tool, the Genesis Commonsense Reasoning and Story Understanding system, and the Analogical Story Merging (ASM) algorithm. They first constructed linkages that allowed stories to be passed from the Story Workbench, through the Commonsense Reasoner, to the ASM system, so that higher-level plot patterns could be discovered automatically and returned to the Genesis system for discovery in an evolving story. They performed an experiment in which they constructed six stories, three of which illustrated the high-level plot pattern of revenge, the other three illustrating the high- level plot pattern of pyrrhic victory. They held on story of each type out for testing, and, using the confederated system, extracted both plot patterns from the remaining four stories. Using the patterns, they then were able to find the correct patterns in the held-out test stories. Also, in preparation for further experiments in this vein, they annotated a corpus of 16 Russian folktales in 16 different representations, approximately 20,000 words, which is the largest, most deeply-annotated narrative corpus assembled to date.