Data Cleaning publication categorizer

In this paper we present a method of parsing unstructured textual records briefly describing a person and their direct relatives, which we use in the construction of a browsing tool for genealogical data. The records have been created by researchers who are currently digitising a collection of historical archives stored at the Abbaye de Saint-Maurice, Switzerland. The string 'Beatrix, daughter of Johannes Trona, of Saillon' is a typical example of a record. We wish to annotate every term (word and symbol) in our records with a label which describes whether the term is a name (e.g. 'Beatrix'), a place (e.g. 'Saillon'), or a relationship (e.g. 'daughter'). Using this information, we are able to derive both a canonical form for each name (e.g. 'Beatrix Trona'), and the relationships between people. We build upon work developed for the cleaning and standardization of names for record linkage corpora, adding several enhancements to deal with our more difficult data, which contains common name structures of French, Italian and Latin, over hundreds of years. We present an approach to this problem that works interactively with a user to annotate the data set accurately, greatly reducing the human effort required. We do this by learning a Hidden Markov Model representing a record structure, and finding structural patterns in new records. Finally, we present a brief overview of a tool we are developing to help genealogical researchers browse and search the data.