MASF: A Multi-Model Adaptive Selection Framework for Abstractive Text summarization

Researchers propose MASF, a Multi-Model Adaptive Selection Framework that combines multiple fine-tuned transformer models with automatic evaluation metrics to improve abstractive text summarization quality. The framework achieves a BERTScore of 88.63% on the CNN/DailyMail dataset, outperforming several large language models including GPT3-D2 and Falcon-7b.

The paper addresses a fundamental limitation in automated text summarization: relying on single models produces inconsistent output quality across diverse article types and topics. MASF introduces an ensemble approach where multiple transformer-based models independently generate candidate summaries for the same input, with automatic evaluation metrics selecting the highest-quality result based on lexical similarity and semantic relevance scores.

This research emerges within a broader trend of moving beyond monolithic AI systems toward adaptive, multi-model architectures. The summarization field has long struggled with the trade-off between specialized models excelling in specific domains versus generalist models offering broader coverage. Ensemble methods have proven effective in other domains, but their application to abstractive summarization with principled selection mechanisms remains relatively nascent.

The framework's practical implications are substantial for content aggregation platforms, news organizations, and information management systems handling high-volume document processing. By achieving superior performance to individual state-of-the-art models, the approach demonstrates that intelligent model selection can extract better results from existing components without requiring larger individual models, potentially reducing computational costs while improving reliability.

Future development directions include investigating how the framework scales with domain-specific corpora beyond news, optimizing the evaluation metrics for different content types, and exploring whether similar adaptive selection strategies could enhance other natural language processing tasks facing quality consistency challenges.

• →MASF combines multiple fine-tuned transformer models with adaptive selection to improve summarization robustness across varied content types.
• →The framework achieved 88.63% BERTScore, exceeding performance of GPT3-D2, Falcon-7b, and other tested large language models.
• →Ensemble approaches with intelligent selection mechanisms can deliver better results than individual state-of-the-art models while potentially reducing computational requirements.
• →The method addresses real inconsistency problems in single-model summarization by leveraging automatic evaluation metrics for quality-based output selection.
• →Results suggest multi-model architectures with adaptive strategies represent a viable path forward for improving reliability in production NLP systems.