<p>Abstractive summarization of dialogues remains a challenging task in natural language processing because conversations are informal, fragmented, and highly context dependent. In this work, we conduct a comprehensive empirical study of an enhanced transformer–extractive hybrid framework that integrates powerful transformer-based architectures (T5 and BART) with lightweight extractive filtering strategies (TF–IDF, TextRank, and a MIX combination). Experiments on the SAMSum dialogue dataset show that fine-tuning substantially boosts performance over zero-shot baselines, and that the hybrid MIX configuration achieves the best trade-off between content relevance and efficiency. T5 consistently outperforms BART across ROUGE and BLEU metrics, particularly under hybrid settings. Ablation analyses quantify the contributions of extractive filtering and fine-tuned encoder–decoder components within the hybrid pipeline. To assess generalizability, we further evaluate the model on an out-of-domain corpus of children’s stories, where it continues to generate coherent and contextually faithful summaries. Evaluation combines ROUGE and BLEU with semantic similarity and factual-consistency-oriented metrics (e.g., BERTScore-based measures), complemented by a small-scale human study and qualitative analysis of strong and weak summaries. Our findings highlight both the strengths and limitations of current filtering and evaluation practices, and provide a practical, scalable framework for bridging extractive and abstractive techniques in dialogue summarization, while outlining remaining opportunities such as more semantic-aware filtering and reinforcement-learning-based optimization.</p>