Story retrieval and comparison using concept patterns

Abstract

To understand a new situation, humans draw from their knowledge of past experiences and events. For a computer to use the same method, it must be able to retrieve stories that shed light on a new situation. Traditional story retrieval uses keywords to determine similarity. Keywords are useful for determining whether stories share similar topics. However, they miss how stories can be structurally similar. In my work, I have used high level concept patterns, which are structures of causally related events. Concept patterns follow the Goldilocks principle, that the features should be of intermediate size. Given a story about cyber crime and another about traditional warfare, the wording will be different, as cyber crime involves viruses, DDOS attacks, and hacking, while traditional warfare involves armies, invasions, and weapons. However, both stories may involve instances of revenge and betrayal. Using a corpus of 15 conflict stories, I have shown that a similarity measure based on concept patterns differs substantially from a similarity measured based on keywords. In addition, I compared three concept-pattern methods with human performance in a pilot study in which 11 participants performed story comparison. My goal was to contribute to a human competence model, but I have also explored applications in story retrieval, prediction, explanation, and grouping.