Universität zu Köln

Brookshire, Patrick Daniel ORCID: 0000-0002-5843-7577 (2026). Tracing Life Events and Their Polarities in 18th Century Biographical Sources with Transformer Models. PhD thesis, Universität zu Köln.

PDF PBrookshire_Tracing-Life-Events.pdf Download (12MB)

Abstract

This thesis shows how life events and their polarities can be traced in (historical) biographies based on natural language processing techniques, such as text classification tasks. To this end, a corpus of 18th century biographical sources is manually annotated with regard to (i) sentiment polarity, (ii) the life stage at which a life event was experienced, and (iii) life event domains. The annotations are used to evaluate the applicability of various systems proposed in related work for similar tasks, and to conduct content analyses of larger samples. This serves three main purposes: First of all, it adresses the question of whether systems trained on modern-day sources are able to adapt to a low-resource setting which is somewhat prototypical for digital humanities projects. In this respect, related work is supported by showing that supervised learning approaches in general and fine-tuned BERT models in particular outperform off-the-shelf systems including prompted large language models. Secondly, this work deals with the field of tension between performance and transparency which arises from the observation that embedding-based language models tend to outperform rule-based systems. This is why explainability approaches are applied in addition to more traditional performance and error analyses. They suggest that the best performing BERT-based sentiment polarity and life event domain classifiers have implicitly learned to (i) recognize relevant textual features and (ii) follow annotation guidelines when labeling ambiguous cases. Finally, this work illustrates how novel categorization schemes derived from sociological and psychological research can be operationalized in a similar way to an established task, such as a sentiment classification. This allows for a change of perspective which offers further analytical options for uncovering less overt patterns and hence more holistic downstream content analyses. One such content analysis suggests that life narratives of girls and women focused more on earlier life stages whereas later parts of life were described in more detail in the case of boys and men. Another one concerns different biographical subgenres and hints at more polarity differences on the narrative side than the narrated one although topical differences are observed in both text and life courses. More findings are discussed in the thesis and related to existing biographical research. In this sense, this work builds up on related one while proposing a methodological approach and describing notable patterns identified with it so that it in turn also opens up many future digital biographical research directions.

Item Type:	Thesis (PhD thesis)
Creators:	Creators Email ORCID ORCID Put Code Brookshire, Patrick Daniel Patrick.Brookshire@adwmainz.de https://orcid.org/0000-0002-5843-7577 UNSPECIFIED
URN:	urn:nbn:de:hbz:38-810452
Date:	2026
Language:	English
Faculty:	Faculty of Arts and Humanities
Divisions:	Faculty of Arts and Humanities > Fächergruppe 1: Kunstgeschichte, Musikwissenschaft, Medienkultur und Theater, Linguistik, Digital Humanities > Institut für Digital Humanities (IDH)
Subjects:	Data processing Computer science Social sciences Language, Linguistics
Uncontrolled Keywords:	Keywords Language Digital Humanities English Natural Language Processing English Text Classification English Sentiment Analysis English Life Stages English Life Events English BERT English Large Language Models English
Date of oral exam:	24 June 2026
Referee:	Name Academic Title Reiter, Nils Prof. Dr. Elwert, Frederik Apl.-Prof. Dr. Cugliana, Elisa Jun.-Prof. Dr.
Refereed:	Yes
URI:	http://kups.ub.uni-koeln.de/id/eprint/81045