How AI can identify the temporal setting of novels

Searching for fiction set in a particular historical period can be difficult. Library catalogues do not always contain detailed information about when the events take place, despite such information being important for both readers and researchers. In her doctoral studies, Fereshta Westin examined how AI-based methods can be used to identify and describe the temporal setting of fiction.

Her research focuses on machine learning and language models that analyse the content of novels to suggest when a story takes place. For example, the models may use historical figures, social conditions, or technological developments as clues to place the events within a particular historical period.

Advanced methods are not always best

One of the most important findings of the thesis is that a relatively simple method, TF-IDF (a statistical method that measures how rare or common a word is within a collection of documents), produced better results than more advanced AI methods when classifying Swedish historical fiction according to the historical period in which it is set.

"Many people expect the most advanced AI methods always to perform best, but my results show that simpler methods can sometimes be more effective in identifying when a story takes place," said Fereshta Westin.

In another part of the project, several language models were asked to suggest start and end years for the events of novels. The results were compared with temporal information in the national library catalogue Libris. The models often proposed more specific time spans than those recorded in the catalogue.

At the same time, the study shows that it is not always obvious how such results should be assessed. A more detailed temporal indication is not necessarily incorrect, but neither is it automatically better than a broader description. The thesis therefore contributes new ways of evaluating AI-generated temporal determinations.

Has developed an analytical framework

The thesis has also resulted in an analytical framework for examining the types of clues language models use when explaining their suggestions. Different types of temporal clues were identified, including years, historical events, people, technology, and social conditions.

"We can see that language models often combine several different types of clues when suggesting a time period. Some are clear, such as a specific year, while others require considerably more interpretation," explained Fereshta Westin.

May provide support for librarians

In the longer term, the results may contribute to the development of digital libraries and information systems. AI could provide librarians with suggestions for temporal settings while also showing which clues in the text underpin the assessment.

"In the future, such tools could support library work and make it easier for both readers and researchers to find literature set in different historical periods," said Fereshta Westin.

"Although the thesis does not develop ready-made solutions, it contributes new knowledge about how AI can be used to organise literature and make it more accessible," said Fereshta Westin.

What is your next step as a researcher?

"I want to continue exploring questions that emerged during my doctoral studies, including how librarians can work with AI-generated suggestions and how stories with multiple timelines can be described in a more nuanced way," she concluded.

Read more

Fereshta Westin defended her doctoral thesis on 9 October in Library and Information Science at The Swedish School of Library and Information Science at the University of Borås.

Read the thesis Temporal subject representation in fiction: From classification to evaluation

Fereshta Westin

Principal supervisor: Professor Gustaf Nelhans, the Swedish School of Library and Information Science

Co-supervisor: Professor Jutta Haider, the Swedish School of Library and Information Science

Read more about research in Library and Information Science

The research is linked to several of the UN Sustainable Development Goals, including Goal 4 on quality education, Goal 9 on industry, innovation and infrastructure, and Goal 16.10 on public access to information.