Public defence of doctoral thesis: Fereshta Westin

Abstact: This doctoral dissertation examines how computational methods can be used to identify the temporal setting of fiction. Temporal setting, referred to in this dissertation as temporal aboutness, concerns the time period in which a fictional work is set and can be understood as one dimension of its subject matter. Despite its potential value for organizing, searching, and retrieving fiction, temporal information is often absent or sparsely represented in library metadata. One reason is that identifying the temporal setting of a novel is not always straightforward. It may be stated directly through dates or historical events, but it may also be conveyed indirectly through language, social conditions, technology, cultural practices, and narrative development. This makes temporal subject analysis both interpretive and time-consuming, particularly across large collections. The dissertation therefore investigates how temporal aboutness can be represented, predicted, evaluated, and interpreted using computational methods.

The dissertation consists of four studies and an overall kappa. The first study reviews previous research on temporal information in text and identifies how time has been conceptualized, represented, and analyzed computationally. The second study examines three methods—TF-IDF, LDA, and SBERT—for classifying Swedish historical fiction into four predefined time periods. The third study uses large language models to predict the temporal setting of fiction as year intervals and evaluates the predictions using both interval overlap and the distance between predicted and reference boundaries. The fourth study analyses explanations generated by large language models to identify the types of textual and contextual cues used to support temporal predictions.

The findings show that temporal aboutness in fiction can be identified computationally, but that it cannot be represented, predicted, or evaluated in a single uniform way. Temporal setting may be represented as a broad historical period, a more precise year interval, or through a combination of explicit and indirect textual cues. These different representations make different aspects of temporal information visible and require different approaches to prediction and evaluation. The results show that temporal patterns can emerge from distinctive words and recurring themes, but also from broader contextual information such as historical references, social conditions, technology, language, and narrative development. Temporal predictions therefore cannot be assessed using a single measure of accuracy, since different representations capture different aspects of temporal alignment. Overall, the findings show that identifying the temporal setting of fiction is a process that ranges from the detection of explicit information to more interpretive forms of inference based on multiple and distributed cues.

The dissertation contributes to research on knowledge organization and subject representation by showing that temporal subject matter in fiction can be represented in several ways: as named time periods, year intervals, and combinations of textual and contextual cues. It further shows that these representations require different computational methods and different forms of evaluation. 

The practical implications are that computational methods can assist catalogers and researchers by identifying candidate periods, intervals, and supporting temporal cues, but their outputs should be reviewed and evaluated in relation to the specific task and representation used.