Modeling social interestingness in conversational stories
Telling stories about our daily lives is one of the most ubiquitous, consequential and seamless ways in which we socialize. Current narrative generation methods mostly require specification of a priori knowledge or comprehensive domain models, which are not generalizable across contexts. Hence, such approaches do not lend themselves well to new and unpredictable domains of observation and interaction, in which social stories usually occur. In this paper, we describe a methodology for categorizing event descriptions as being socially interesting. The event sequences are drawn from crowd-sourced Plot Graphs. The models include lowlevel natural language and higher-level features. The results from classification and regression tasks look promising overall, indicating that general metrics of social interestingness of stories could be modeled for sociable agents.
