Besides the importance of events and the arrangement of events, the driver of the events, or the character of the story who acts out of their motivations, is another essential research object in the field of Narratology. The present study tries to come up with an automatic analysis method, namely agency analysis, which can tell us how characters’ actions can reveal the extent to which they intend to interact with their surrounding environment, forming a complementary and easy-to-interpret research method for the field of computational narratology.

We design our way to analyze the agency of fictional characters by adopting a specific linguistic feature - the thematic roles that a character plays in a narrative, hoping that this syntax-semantics interface, which indicates the predicate-argument relationship in an entity-based manner, can reveal meaningful insights into a character’s intentionality on a psychological level. By including the agency of fictional characters in computational literary studies, we will have an opportunity to associate fictional characters’ actions in an imaginary world with the complex dynamics of human decision-making and behavior, and help us gain a deeper understanding of the complex relationships between characters, authors, and readers, or develop more nuanced and sophisticated theories of narrative and storytelling.

In general, thematic roles refer to “the roles that participants play in the events described by verbs” (McRae et al., 1997), which involves semantic interpretations of every syntactic component of the sentence (e.g., some roles are decided by the nature of the verb). Linguists have categorized the roles as Agent, Patient, Goal, Experiencer, Stimulus, etc. (Dowty, 1991). However, it is hard to provide precise definitions of these thematic roles and even harder to build an automatic pipeline that differentiates and labels the “clusters of concepts” (Dowty, 1991) accurately.

In that case, we will stick to the annotation scheme of PropBank (Babko-Malaya, 2005) and the notions of Proto-agent and Proto-patient introduced by Dowty (1991) to investigate the distribution of proto-roles played by a fictional character, and use AllenNLP (Gardner et al., 2018), a deep-learning-based Python toolkit, to annotate whether a protagonist plays the role of Pro-Agent or Pro-Patient in a sentence, as they are the two most relevant roles to agency and carry the information of actions. By adopting proto-roles instead of the traditional classification, such as Agent, Patient, Theme, and Experiencer, we can better focus on how verb-based events reflect a fictional character's agentive behavior in general and avoid the possible ambiguities caused by a complicated categorization methodology.

To better understand what agency analysis can accomplish with large-scale textual analysis and test the capability of our measurement to complement previously used computational methods, such as sentiment analysis, we construct a protagonist-centric dataset comprising 300 protagonists from 261 books sourced from Project Gutenberg. By calculating each character’s level of agency and sentiment, we identify a shift in authors' writing preferences over time, with more recent books depicting more agentive yet less happy characters. Furthermore, we observe gender and narrative point-of-view biases, with female characters depicted with less agency than male characters and first-person narrators exhibiting higher levels of agency. Our approach offers a complementary perspective for contemporary computational literary studies and will help to reveal how different features of characters interact and influence each other within the storytelling process.

Bibliography:

Dowty, D.R. (1991). Thematic proto-roles and argument selection. Language, 67, 547–619.

McRae, K., & Ferretti and Liane Amyote, T. R. (1997). Thematic roles as verb-specific concepts. Language and cognitive processes, 12(2-3), 137-176.

Babko-Malaya, O. (2005). Propbank annotation guidelines. URL: http://verbs.colorado. edu.

Gardner, M., Grus, J., Neumann, M., Tafjord, O., Dasigi, P., Liu, N., ... & Zettlemoyer, L. (2018). Allennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640.