Artificial empathy : using vector space modeling and mixed scope alignment to infer emotional states of characters in stories

Abstract

Emotions greatly influence human cognition. Therefore, if we are to develop artificially intelligent programs that work closely with humans, we must ensure that they are capable of empathy. In an effort to realize the goal of emotionally aware programs, I created a multi-corpus informed vector space model to determine the emotions evoked by individual terms. I then combined that information with the semantic parse trees produced by the Genesis Story Understanding System to ascertain the emotions evoked by a single sentence. Additionally, I used the story aligner within Genesis to determine the emotions evoked by stories described over multiple sentences. My program can infer characters' emotional states based on their descriptions, the situations they are involved in, and the actions they perform. For instance, it infers that Alice is joyful from the sentence "Alice wins an award" and that James is probably experiencing sadness from the sentence "James is lonely." Additionally, the program can identify that Austin is likely surprised if "Austin has to take a test" and "Austin doesn't know about the test."