Machine Learning vs Deterministic Rule-Based System for Document Stream Segmentation

Classical document stream Segmentation methods rely on physical separators (white pages, pages with a specific stamp, etc) to automatically split documents from the stream (detecting the beginning and the ending of documents). In order to reduce costly efforts, a recent work using a contextual rulebased approach was proposed to automate this process. Such rules tend to detect continuity, rupture or uncertainty between pairs of pages. Even if these first results were encouraging, performance remained unsatisfactory. In this context, we propose to compare this existing rule-based approach to a machine learningmethod basedon Doc2Vecsoas toevaluate andcompare their strengths and weaknesses. This study was led on a corpus of more than 4,000 real administrative documents composed of more than 8,000 pages. The machine learning approach gives better results on multipage documents while the rule-based method performs best with single page documents.