An Ensemble Framework for Dynamic Character Relationship Sentiment in Fiction

Abstract

AbstractFictional characters in a narrative text can experience various events in the narrative timeline as the progress of character development. Relationships between characters can also dynamically change over time. Summarizing the relationship dynamics in fiction through manual annotation can be very tedious even at a small scale, but highly impractical or even impossible in a large corpus. With the recent development of machine learning models in Natural Language Processing, many tasks have been introduced to help humans extract information from text automatically. Motivated by this development, we propose a conceptual model and an information extraction framework that combines two state-of-the-art machine learning algorithms to extract character relationships directly from an event sentence in a fictional narrative. For our use case, as we consider sequence in a story line, we also infer the dynamic sentiment relationships among characters over time. Since this approach is by nature unsupervised, we also preserve the provenance of each relation extracted in order to prepare a dataset to use in training a supervised model. We hope this approach can be a step toward more robust automatic character relation and event extraction from fictional texts.KeywordsDigital libraryInformation extractionMachine learningNetwork analysis