Extracting economic narratives using natural language processing
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Alternative Title(s)
A temporal perspective
Abstract
As our society has become increasingly digitized across the last decades, the need to analyze unstructured data sources has grown in many sectors. In particular, economic and political researchers have been interested in quantitatively uncovering information found in texts, as theoretical research in these fields points towards the great effects of narratives on economic decision making. While such economic narratives can follow a myriad of definitions, they can generally be described as sense-making stories that can influence the reader's economic or political decisions in the future. Developing quantitative methods that can extract and analyze narratives from texts is therefore a major step to better understand market behavior or the decisions of policy makers, among others.
In this cumulative dissertation, I outline the research I have conducted concerning economic and political narratives with a particular focus on using diachronic language modeling that enables me to analyze narratives not in a vacuum, but rather observe narrative shifts over time. To do this, I first properly define economic narratives and provide an overview of language models I used in my research. I then proceed to summarize the methodology and contributions to the current research proposed in my papers, starting with works that do not consider a temporal component when extracting narratives from texts. These works show the development of narrative extraction techniques over time, starting from an unsupervised text classification method and a large pipeline of models specifically designed to handle the task, and ending with the use of state-of-the-art Large Language Models to extract narratives utilizing their great language understanding capabilities.
After covering atemporal narrative extraction methods, I focus on my works that utilize diachronic language modeling. I present two diachronic change detection methods, designed to identify points in time at which we can suspect a narrative shift. The first method uses the frequency of mentions of an entity in the media over time to detect a change, enabling an analysis of narratives surrounding that entity. The second method detects changes in the topics of the topic model LDA, allowing for an analysis of the corpus at large rather than a specific entity. I then propose a method to combine temporal and atemporal methods, with atemporal methods extracting narratives (or narrative shifts) and the detected change points. I further present software publications that contain all diachronic methods proposed in my works. Lastly, I conclude by giving an outlook on future research.
Description
Table of contents
Keywords
Natural language processing, Economic narratives, Diachronic language models, Python
Subjects based on RSWK
Narrativität, Wirtschaftswissenschaften, Empirische Forschung, Natürlichsprachiges System