Entropy-guided graph-based ensemble classification for legal rhetorical role labeling
摘要
Rhetorical Role Labeling (RRL) in legal judgment documents is challenging due to feature redundancy, and the constraints of individual classifiers. The task becomes more efficient and effective when only the most significant features are considered during learning of classifiers. Graph-based ensemble learning further amplifies the benefits of feature optimization by leveraging complementarities among classifiers, resulting in a more stable and generalizable model. We propose a two-stage framework that performs feature selection followed by ensemble learning. As an initial step, conditional entropy-based feature subset selection is devised to remove noisy and redundant features while retaining the most discriminative ones. Next, a graph-based ensemble model is learned on weighted directed graph constructed using reduced dataset. In the generated graph, each node represents a base classifier trained on the optimal feature subset, and a directed edge is considered from a node to another node if and only if similarity of first node on second one in the pair is high ensuring a generalized graph-based ensemble model. Classifier importance within the ensemble is assessed using graph metrics that capture both local and global aspects. To enhance diversity and mitigate individual classifier weaknesses, classifier similarity is modeled using F1-score embeddings, and diverse ensembles are extracted via Minimum Spanning Trees (MSTs) with the Chu–Liu/Edmonds’ algorithm. On the UK’s House of Lords judgments (HOLJ) corpus, the original feature set {