<p>Automatic text summarization and evaluation have become an important research topic in the field of natural language processing (NLP), especially in the era of rapidly increasing digital information. Generating concise and informative summaries from large volumes of textual data remains a challenging problem, particularly in multi-document environments where identifying relevant content and preserving contextual relationships is complex. Many existing studies have explored issues related to datasets, evaluation metrics, and model architectures; however, significant challenges still remain in effectively capturing contextual nuances and maintaining coherence in generated summaries. To address these limitations, this work introduces a unified deep learning-based framework for multi-dimensional evaluation of summaries called <i>EvaS</i>. The proposed <i>EvaS</i> model evaluates summaries using four important quality dimensions such as fluency, factual consistency, coherence, and relevance. Each dimension is assessed independently using neural models built upon pre-trained language representations from BERT and GPT-2. The individual dimension scores are then combined using a harmonic mean to obtain an overall evaluation score. Based on the aggregated score, the framework ranks candidate sentences according to their importance, enabling the selection of content that best represents the key information in the source text. This mechanism supports the development of concise abstractive summaries while preserving the essential meaning and contextual relationships of the original documents. The performance of <i>EvaS</i> is evaluated using various baseline metrics, and the experimental results demonstrate promising improvements compared with several baseline approaches. In particular, the proposed method achieves an overall aggregate correlation score of 0.65, indicating its effectiveness in assessing summary quality across multiple dimensions.</p>