Summarization of Scientific Articles using Hybrid SciBERT and Graph-Based Algorithms

In the ever-evolving information age, automatic text summarization is an important solution to address complex text processing problems, especially in scientific articles. Scientific articles often have complex structures and terminology, requiring methods to capture deep contextual meaning. This research proposes an extractive text summarization approach that combines SciBERT, a BERT-based language model developed specifically for scientific texts, with PageRank and HITS algorithms that are integrated using Harmonic Mean. SciBERT converts sentences in scientific articles into sentence vectors by combining the last four layers to produce optimal contextual embedding to preserve complex scientific meaning and context. Furthermore, PageRank and HITS algorithms rank sentences based on relevance and the relationship between sentences. Combining these two algorithms through the Harmonic Mean produces a more representative summary, balancing sentence relevance and overall text coherence. The approach was tested on the PubMed dataset, showing that the proposed model could produce summaries with ROUGE-1 scores reaching 32.23%, ROUGE-2 reaching 11.64%, and ROUGE-L reaching 29.18%. These results show that combining SciBERT with graph-based algorithms can effectively capture and retain important information in complex scientific texts while overcoming problems such as loss of context between sentences.