In this thesis we propose methods to model inter-personal relationships in text. Due to their inherent social nature, people continuously interact and form relationships with each other.
Understanding these relationships is essential to understanding and explaining people's desires, goals, actions and expected behaviors. Apart from applications related to general natural language understanding, modeling inter-personal relationships also finds application in many real-world domains such as social networks, discussion forums etc.
In this proposal we provide methods to model inter-personal relationships in natural language text with a focus on narratives. We demonstrate that such a task can benefit from using models that are capable of incorporating not just linguistic cues but also the contexts in which these cues appear.
We consider two types of narratives: movies and novels, and propose structured models to address the task of modeling the nature of relationships between any two given characters from the narrative. We
attempt to jointly infer the nature of relationships between all characters in the narrative and demonstrate how the task of identifying relationship between two characters can benefit by including information about their
relationships with other characters in the movie. We next formulate the relationship-modeling problem as a structured prediction task to acknowledge the evolving nature of human relationships and demonstrate the need to model history of relationships between characters while modeling their current relationship. We then propose to jointly address the relationship-modeling task in the two domains mentioned above to better utilize their commonalities while simultaneously teasing apart their idiosyncrasies.
Lastly, we demonstrate a practical application of this task. We analyze contents of online educational discussion forums to automatically suggest threads to the instructors that require their intervention. By suggesting avenues for instructor-student interaction, we alleviate the need for the instructor to manually go over all threads of the forum and also help the students who have no way of interacting with the instructor. We propose to incorporate thread structure into our approach by using latent variables that abstractly represent contents of individual posts and model the flow of information in the thread.
Committee Chair: Dr. Hal Daume III
Dept's Representative Dr. Hector Corrada Bravo
Committee Member(s): Dr. Chris Dyer (CMU, LTI)
Dr. Philip Resnik