View article

In this paper we present an approach for the automatic extraction of family relationships from a real-world collection of historical notary acts. We retrieve relationships such as husband-wife, parent-child, widow of, etc. We study two ways to deal with this problem. In our first approach, we identify all person names in a document, generate all potential candidate pairs of names and predict whether they are related to each other using classification techniques where the text fragments that occur around and between two names are sued as features.

In the second approach, we train and apply a Hidden Markov Model to annotate every word in a document with an appropriate tag indicating if it is a name, a specified relationship descriptor, or neither of these. Then we look for the names connected to each other via relationship descriptors. We discuss the challenges such as processing raw data, obtaining a sufficient amount of training examples, and dealing with an imbalanced and noisy collection. We evaluate our results for each relationship type in terms of precision, recall and f-score.