Life Trajectories as Symbolic Language

Deep learning has significantly advanced research within natural language processing in recent years. Beyond language, novel transformerbased architectures have shown promise as tools to model various multivariate sequences. These include weather patterns, musical compositions, and protein structures. Similarly, human lives represent another form of multivariate sequences comprising various events: People are born, attend kindergarten, visit doctors, relocate to new cities, and more.
By drawing parallels between human lives and written language, we propose a novel methodology to study individual life trajectories. We use the Danish National Registry to create an artificial symbolic language. It transforms socioeconomic and health events into a structured, sentence-like format, akin to the words and sentences in a language.
This representation approach lays the groundwork for our primary contribution: developing the life2vec model, a transformer-based model designed for analyzing life trajectories. In the thesis, we demonstrate that the life2vec model captures complex relationships between various life events and uses this knowledge to provide insights into early mortality and personality. A key strength of life2vec lies in interpretability, as we can use it to explore the influence of socioeconomic and health factors on individual life paths.
The findings from this thesis underscore the potential of transformermodels in understanding and predicting human behavior and experiences.