Boundary-Centered Explainability for Dialogue Topic Segmentation

Despite recent advances in dialogue topic segmentation, existing work provides limited evidence about why individual utterances are predicted as boundaries and how local explanations depend on the selection strategy and perturbation protocol. Using fixed checkpoints of 3LHSeg, a hierarchical dialogue topic segmentation model, we evaluate a boundary-centered framework that compares five local utterance-selection strategies through complementary perturbation diagnostics. The audit also examines configuration sensitivity and explanation stability and uses Random-Baseline Gain (RBG) as a diagnostic of relative local distinctiveness by contextualizing comprehensiveness against matched random subsets from the same local candidate window. Experiments on TIAGE, QMSum, and Friends show that Leave-One-Out ranks most favorably under the adopted perturbation-based diagnostics, although this result is protocol-specific. The diagnostics provide partially overlapping information and vary with the local configuration. Raw deletion area-under-the-curve values are consistently positively associated with initial boundary confidence, but a centered control substantially attenuates this association in most dataset–strategy combinations. Overall, explanation behavior, probability quality, confidence, stability, and boundary correctness should be treated as distinct dimensions, supporting multi-diagnostic and configuration-explicit auditing rather than reliance on a single explanation score.