Named Entity Recognition in the Historical Meeting Protocols of the Tartu City Council

In this paper, we describe our efforts to enable automatic named entity recognition in the historical meeting protocols of the Tartu City Council from 1918–1940, written in Estonian. We describe the conversion of automatically transcribed protocol pages into a form suitable for manual corpus annotation and model building, introduce our choice of name categories, and describe the manual annotation process. The resulting corpus is annotated for 8 named entity categories, with fairly good inter-annotator agreement on the majority of the categories. Finally, we present experiments on fine-tuning two BERT-like models on the dataset, and report that the best model achieved an overall F1-score of 0.8246.