Summarization of Legal Texts with High Cohesion and Automatic Compression Rate

We describe a method for extractive summarization of legal judgments using our own graph-based summarization algorithm. In contrast to the connected and undirected graphs of previous work, we construct directed and disconnected graphs (a set of connected graphs) for each document, where each connected graph indicates a cluster that shares one topic in a document. Our method automatically chooses the number of representative sentences with coherence for summarization, and we don’t need to provide a priori, the desired compression rate. We also propose our own node/edge-weighting scheme in the graph. Furthermore, we do not depend on expensive hand-crafted linguistic features or resources. Our experimental results show our method outperforms previous clustering-based methods, including those which use TF*IDF-based and centroid-based sentence selection. Our experimental results also show that our method outperforms previous machine learning methods that exploit a variety of linguistic features.