Investigating the Use of Word Embeddings to Estimate Cognitive Interest in Stories
- Morteza Behrooz, Computer Science / Computational Media, University of California Santa Cruz, Santa Cruz, California, United States
- Justus Robertson, North Carolina State University, Raleigh, North Carolina, United States
- Arnav Jhala, North Carolina State University, Raleigh, North Carolina, United States
AbstractNarrative processing is an important skill to model both from a cognitive science perspective and a computational modeling perspective which applies to intelligent agents. Communication between humans often involves storytelling patterns that make the mundane exchange of information more interesting and with proper emphasis on important communicative goals. Current narrative generation models evaluate their generations based on either a priori domain semantics (e.g. game state for an in-game conversation with player agents) or generic text quality measures (e.g. coherence). However, in utilizing storytelling as a communicative tool for real-world interactions, domain-specific approaches fail to generalize and text quality measures fail to ensure that the narrative is perceived as interesting. Hence, such generation needs to consider the cognitive processes involved in the perception of narrative. Using theories of cognitive interest, we present results of an investigation of whether word embeddings (e.g. GloVe (Pennington,Socher, & Manning, 2014)) could be used to model and estimate cognitive interestingness in stories.
Return to previous page