Automatic Identification of Relations in Quebec Heritage Data

摘要

Heritage data is often represented in unstructured format, especially textual data. In this paper, our objective is to extract instances of predefined relations between persons and real estates from historical notices in French. Using several vector-based representations and supervised learning algorithms, we build classifiers able to achieve an F-measure between 75% to 85% for relation detection. Our results show that performances are highly dependent on the type of relation, and also on the specific evaluation metrics. Our best results are obtained using a TF-IDF vector representation with a support vector machine classifier or Word2Vec vectors combined with a multilayer perceptron classifier.