Coherence Graphs: Bridging the Gap in Text Segmentation with Unsupervised Learning

In the domain of Text Segmentation (TS), substantial advancements have been achieved through the introduction of innovative supervised learning methodologies. Recent years have witnessed supervised methods surpassing traditional unsupervised techniques in effectiveness. However, this enhanced performance is accompanied by the prerequisites of extensive training datasets and significant initial training duration. Our research relies on an unsupervised approach to TS by devising a graph-based keyword storage mechanism. This architecture results in performance improvements on a variety of datasets in comparison to current state-of-the-art unsupervised TS approaches, as evaluated by \(P_k\) and WindowDiff metrics. Furthermore, this study delves into the application of contemporary large-scale Language Large Models (LLMs), including GPT-4, for executing TS tasks.