Article: Efficient post-training pruning of dialogue summarisation based on sensitivity evaluation Journal: International Journal of Applied Pattern Recognition (IJAPR) 2026 Vol.8 No.3 pp.1 - 18 Abstract: We propose an efficient post-training structured pruning framework for dialogue summarisation that directly optimises under hardware FLOPs constraints, eliminating the need for full-model retraining. Sensitivity of attention heads and FFN filters is estimated via the Hessian matrix trace approximated by the Hutchinson algorithm, guiding a FLOPs-constrained mask search followed by intra-layer greedy rearrangement. A layer-wise Huber regression step then fine-tunes binary masks to real values to restore output fidelity. Experiments on five benchmark datasets show that the framework reduces FLOPs by 30%-40% with less than 1% degradation across BLEU, ROUGE, METEOR, and PARENT metrics, achieving up to 1.52× inference acceleration. Inderscience Publishers - linking academia, business and industry through research

Title: Efficient post-training pruning of dialogue summarisation based on sensitivity evaluation

Authors: Yupeng Liu; Yuhao Zhang; He Sun; Xiaochen Zhang

Addresses: Harbin University of Science and Technology, No. 52 Xuefu Road, Nangang District, Harbin 150080, Heilongjiang Province, China ' Harbin University of Science and Technology, No. 52 Xuefu Road, Nangang District, Harbin 150080, Heilongjiang Province, China ' Harbin University of Science and Technology, No. 52 Xuefu Road, Nangang District, Harbin 150080, Heilongjiang Province, China ' Heilongjiang Institute of Technology, No. 999 Hongqi Street, Daowai District, Harbin 150050, Heilongjiang Province, China

Abstract: We propose an efficient post-training structured pruning framework for dialogue summarisation that directly optimises under hardware FLOPs constraints, eliminating the need for full-model retraining. Sensitivity of attention heads and FFN filters is estimated via the Hessian matrix trace approximated by the Hutchinson algorithm, guiding a FLOPs-constrained mask search followed by intra-layer greedy rearrangement. A layer-wise Huber regression step then fine-tunes binary masks to real values to restore output fidelity. Experiments on five benchmark datasets show that the framework reduces FLOPs by 30%-40% with less than 1% degradation across BLEU, ROUGE, METEOR, and PARENT metrics, achieving up to 1.52× inference acceleration.

Keywords: sensitivity evaluation; post-training pruning; FLOPs-constrained pruning; dialogue summarisation.

DOI: 10.1504/IJAPR.2026.154810

International Journal of Applied Pattern Recognition, 2026 Vol.8 No.3, pp.1 - 18

Received: 22 Nov 2025
Accepted: 16 Apr 2026
Published online: 14 Jul 2026 *

Title: Efficient post-training pruning of dialogue summarisation based on sensitivity evaluation

Keep up-to-date