The rapid growth of textual content has introduced a significant challenge in efficiently extracting targeted information from long documents such as books. This highlights the need for advanced techniques capable of summarizing large volumes of data accurately and effectively. However, the existing summarization approaches tend to produce generic summaries that overlook specific topics or user needs, limiting their effectiveness for personalized or topic-specific purposes.

This research presents Topic-Aware Abstractive Text Summarization (TAATS), a new dual encoder–decoder model designed to generate concise, coherent, and topic-focused summaries. TAATS is built upon the Longformer Encoder–Decoder architecture and incorporates a dual-input design,

where both the source text and a refined topic are encoded in parallel and fused using a multihead attention mechanism. An entity-aware auxiliary loss is also introduced to enhance factual consistency and improve the alignment between the generated summary and the guiding topic.

To support the development and evaluation of TAATS, a new benchmark dataset, TOPSUMSci,

was constructed. It comprises 1,000 annotated records extracted from 495 scientific survey

papers across 33 computer science domains. Each record includes a long paragraph, five refined

topics derived using TF-IDF and Latent Dirichlet Allocation, and five corresponding topic-guided summaries.

TAATS was evaluated against three alternative summarization methods, demonstrating up to

43% improvement in ROUGE-L, 37% in ROUGE-2, and a 37% in BLEURT, confirming its effectiveness

in generating topic-relevant summaries. Semantic evaluations using BERT-based metrics

further confirmed its superior performance in focus, conciseness, and high-level coherence. Although scientific texts served as the case study, the proposed method is generalizable to other domains involving topic-conditioned summarization tasks.

​​

​​