Could you elaborate on the specific resource-limited scenarios (e.g., edge devices, real-time text summarization systems) that your work primarily targets, and what are the key computational/energy constraints in these scenarios that make large PLMs infeasible?
This work targets resource-constrained summarization scenarios where both computational resources and domain-specific training data are limited. A representative example is edge-device applications such as e-book readers with built-in automatic summarization functions. In such products, documents span a wide range of topics and genres, making it impractical to fine-tune a separate language model for each domain. Moreover, e-book readers and similar edge devices typically rely on low-power CPUs and battery-based operation, where repeated PLM inference would introduce unacceptable latency and energy consumption.
In addition, many real-world applications involve emerging or specialized domains where annotated summarization data is scarce or unavailable. In these settings, PLM-based methods often suffer from degraded performance due to domain mismatch and the lack of fine-tuning data. In contrast, the proposed framework does not rely on domain-specific supervision. Instead, it formulates extractive summarization as a multi-objective optimization problem that focuses on structural properties such as coverage, redundancy control, and length constraints, which remain stable across domains. The two-stage design further enhances adaptability by separating sentence selection from semantic refinement: NSGA-II generates structurally valid candidate summaries, while pre-trained language models are used only as auxiliary semantic scorers. As a result, the framework maintains stable performance across different datasets and application scenarios, making it particularly suitable for low-resource, low-power, and data-scarce environments.
