Automatic Detection of Core Sections in Scientific Papers

Abstract Academic research workflows can be significantly accelerated by tools that provide targeted access to the most relevant parts of scientific articles, avoiding the need to read full documents. Since interest typically centres on key elements such as the research problem, proposed approach, and main findings, automatic detection of a paper’s structural sections can greatly facilitate navigation and content retrieval. In this work, we address the task of sentence-level section classification in research papers, assigning each sentence to categories such as Introduction , Related Work , Proposed Approach, Body, Results, and Conclusion . We propose a classification pipeline built on a high-quality dataset constructed via a dedicated filtering and selection process. Two pre-trained models are first used to automatically label sentences extracted from research articles; only sentences for which both models agree on the predicted label are retained, ensuring consistent annotation. On this refined dataset, we fine-tune a BERT-Base* classifier that achieves an average precision of 84.67%, recall of 83.78%, F1 score of 84.86%, and an overall accuracy of 86.49%, indicating robust and well-balanced performance across classes. The dataset, model, and implementation code are publicly available. By enabling accurate sentence-level structural tagging, our approach supports more efficient research workflows and lays the groundwork for advanced applications in summarisation, information retrieval, and semantic search over large collections of scientific documents.