Processing the structure of documents: Logical Layout Analysis of historical newspapers in French

Background. In recent years, libraries and archives led importantdigitisation campaigns that opened the access to vast collections of historicaldocuments. While such documents are often available as XML ALTO documents, theylack information about their logical structure. In this paper, we address theproblem of Logical Layout Analysis applied to historical documents in French.We propose a rule-based method, that we evaluate and compare with twoMachine-Learning models, namely RIPPER and Gradient Boosting. Our data setcontains French newspapers, periodicals and magazines, published in the firsthalf of the twentieth century in the Franche-Comt\'e Region. Results. Ourrule-based system outperforms the two other models in nearly all evaluations.It has especially better Recall results, indicating that our system covers moretypes of every logical label than the other two models. When comparing RIPPERwith Gradient Boosting, we can observe that Gradient Boosting has betterPrecision scores but RIPPER has better Recall scores. Conclusions. Theevaluation shows that our system outperforms the two Machine Learning models,and provides significantly higher Recall. It also confirms that our system canbe used to produce annotated data sets that are large enough to envisageMachine Learning or Deep Learning approaches for the task of Logical LayoutAnalysis. Combining rules and Machine Learning models into hybrid systems couldpotentially provide even better performances. Furthermore, as the layout inhistorical documents evolves rapidly, one possible solution to overcome thisproblem would be to apply Rule Learning algorithms to bootstrap rule setsadapted to different publication periods.Comment: 25 pages, 5 figures, 16 tables. This paper is to be submited to the NLP4DH edition of the Journal of Data Mining and Digital Humanitie