Learning narrative structure from annotated folktales

摘要: Narrative structure is an ubiquitous and intriguing phenomenon. By virtue of we recognize the presence Villainy or Revenge in a story, even if that word not actually present text. anvil for forging new artificial intelligence machine learning techniques, window into abstraction conceptual as well culture its influence on cognition. I advance our understanding narrative by describing Analogical Story-Merging (ASM), algorithm can extract culturally-relevant plot patterns from sets folktales. demonstrate ASM learn substantive portion Vladimir Propp's influential theory folktale plots. The challenge was to take descriptions at one semantic level, namely, event timeline described folktales, abstract next higher level: structures such Villainy, Struggle-Victory, Reward. based Bayesian Model Merging, technique regular grammars. that, despite ASM's large search space, carefully-tuned prior allows converge, furthermore it reproduces categories with chance-adjusted Rand index 0.511 0.714. Three important are identified F-measures above 0.8. data 15 Russian comprising 18,862 words. a. subset original tales. This annotated 18 aspects meaning 12 annotators using Story Workbench, general text-annotation tool developed this work. Each aspect doubly-annotated adjudicated inter-annotator cluster around 0.7 It largest, most deeply-annotated corpus assembled date. work has significance far beyond First, points way toward applications many domains, including information retrieval, persuasion negotiation, natural language generation, computational creativity. Second, semantics skill underlies cognitive tasks, so provides insight those processes. Finally, opens door cultural influences cognition differences captured stories.