GPU-accelerated extractive multi-document text summarization using decomposition-based multi-objective differential evolution

Muhammad Hafizul Hazmi Wahab,Nor Asilah Wati Abdul Hamid,Shamala Subramaniam,Rohaya Latip,Mohamed Othman
DOI: https://doi.org/10.1016/j.eswa.2024.125951
IF: 8.5
2024-12-11
Expert Systems with Applications
Abstract:Multi-document text summarization is computationally intensive, mainly when employing complex optimization algorithms. The computational demands increase significantly due to the integration of complex optimization algorithms and the computationally expensive repair operator. As the complexity of the optimization process grows and larger datasets are processed, execution time becomes a critical bottleneck, making real-time summarization challenging. A novel approach, decomposition-based multi-objective differential evolution (MODE/D), was introduced to address these demands in a serial execution context, but its sequential design leads to long execution times when applied to large datasets. This paper introduces the first-ever GPU-accelerated decomposition-based multi-objective differential evolution (GMODE/D), specifically designed to overcome these performance bottlenecks in optimization-based extractive multi-document text summarization. The proposed GMODE/D algorithm introduces two novel execution variants: Variant I, where the enhanced sentence scoring repair operator is executed on the CPU, and Variant II, where the sentence scoring is offloaded to the GPU to enhance performance further. These variants enable the exploration of computational models that balance CPU and GPU tasks in heterogenous environments. Experiments conducted on Document Understanding Conferences (DUC) datasets demonstrate that GMODE/D achieves a speedup of 18.17× over MODE/D and processes summaries at a rate of 215.52 words per second (WPS). Additionally, GMODE/D maintains high summary quality, achieving notable ROUGE-1, ROUGE-2, and ROUGE-L scores. The results show that GMODE/D significantly reduces execution time, setting a new benchmark in the performance of optimization-based extractive text summarization approaches.
computer science, artificial intelligence,engineering, electrical & electronic,operations research & management science
What problem does this paper attempt to address?