Advancing text summarization with specialized datasets: computer science and geography domains

Abstract Automatic text summarization (ATS) has seen considerable development, but state-of-the-art models depend on vast training data, limiting their applicability in scholarly communication and digital knowledge management for technical and domain-specific texts where annotated datasets are scarce. This article addresses this critical gap. First, to create a controlled, data-constrained setting for study, we created two specialized datasets: CS-Summ and Geog-Summ, comprising 794 and 438 manually summarized excerpts from the Computer Science and Geography domains, respectively. We then propose and evaluate a simple, computationally inexpensive method of keyword-guided fine-tuning to enhance performance. By prepending extracted keywords, identified using spaCy’s NER model, to the source document as an explicit signal of salience, we guide the model’s attention during generation. Using these datasets, we conduct baseline experiments by keyword-guided fine-tuning on T5, BART, and BERT models and comparing their performance against zero-shot GPT-4. Evaluation using standard lexical metrics demonstrates that our keyword-guided approach consistently improves performance over the fine-tuned baselines. More critically, using factual consistency metrics such as BERTScore and NLI-based metrics for abstractive summarization, we show that keyword guidance reduces model hallucination and produces more reliable and factually accurate summaries. Our findings advance computational methods for technical content summarization and underscore the practical value of this method in the development of more robust and accessible summarization systems in resource-constrained environments, contributing to broader efforts to develop digital tools that support specialized scholarly analysis.