Multimodal Character Representation for Visual Story Understanding

Stories are one of the main tools that humans use to make sense of the world around them. This ability is conjectured to be uniquely human, and concepts of agency and interaction have been found to develop during childhood. However, state-of-the-art artificial intelligence models still find it very challenging to represent or understand such information about the world. Over the past few years, there has been a lot of research into building systems that can understand the contents of images, videos, and text. Despite several advances made, computers still struggle to understand high-level discourse structures or how visuals and language are organized to tell a coherent story.

Recently, several efforts have been made towards building story understanding benchmarks. As characters are the key component around which the story events unfold, character representations are crucial for deep story understanding such as their names, appearances, and relations to other characters. As a step towards endowing systems with a richer understanding of characters in a given narrative, this thesis develops new techniques that rely on the vision, audio and language channels to address three important challenges: i) speaker recognition and identification, ii) character representation and embedding, and iii) temporal modeling of character relations. We show that our approach improves systems ability to understand narratives, which is measured using several tasks such as their ability to answer questions about stories on several benchmarks.