Year: 2026
Type: conference-paper
Abstract: In this paper, we present a series of experiments using historical language models to investigate the impact of pretraining on data that more closely resembles the task domain, focusing on the case study of automatic olfactory event extraction.We tested five monolingual historical and contemporary pretrained models on the task of extracting olfactory events using a benchmark spanning several centuries.The languages investigated are English, French, Italian, German, and Dutch.The aim of our research is not only to assess whether historical models can improve performance on this diachronically oriented task, but also to gain deeper insight into the factors influencing model performance through a detailed analysis of performance patterns.We examine potential sources of variation and previously proposed hypotheses to account for lower performance observed in this task, thereby offering a more comprehensive understanding of model behavior in this context.
Authors: Teresa Paccosi, Marijn Koolen
Institution: KNAW Humanities Cluster
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166567987