Collaborative Multi-agent System for Automatic Linear Text Segmentation

This paper proposes a collaborative multi-agent system for splitting documents into semantically coherent text chunks, labeling them according to a given segmentation structure. Diverse linear text segmentation methods can be incorporated into the system by introducing new agents, which allows to combine complementary approaches: domain-specific, supervised and unsupervised. The system must be supplied with a representative set of previously segmented documents from the target corpus, which are used both to train the supervised agents and to evaluate every agent within the system, similar to ensemble methods. The accuracy of each agent determines its weight in a subsequent aggregation phase, when a common solution is agreed on. The proposed approach presented promising results on segmenting documents from a juridical corpus.