Hybrid abstractive model using BioBERT and stacked LSTM for biomedical text summarization

In the rapidly evolving field of biomedical research, the ability to efficiently and accurately summarize vast amounts of text is essential. Hallucination in the context of summarization refers to the generation of information that is not present in the source text. Generating an abstractive summary poses a significant challenge, especially in domains like biomedicine, where it’s difficult to balance the accuracy with other summary characteristics, such as readability and contextual understanding. This paper proposes a HE2A_BTS model for generating abstractive summaries of biomedical research articles that balances factual accuracy with linguistic clarity. The model utilizes BioBERT for extractive summarization in the first phase, followed by a stacked LSTM with a UMLS (Unified Medical Language System) based Attention layer and copying mechanism for abstractive summarization in the second phase. The performance evaluation of the proposed hybrid abstractive summarization model is based on multidimensional analysis, which combines both automated evaluation metrics and human-expert scoring. These include BERTScore (78.07) for semantic alignment, ROUGE (60.10) for n-gram and structural overlap, and simulated FactCC (83.5) for factual consistency. The comprehensive evaluation indicates that HE2A_BTS is effective in generating semantically faithful, readable, and factually consistent biomedical summaries that retains the complex biomedical terms. By balancing linguistic simplification with information integrity validated through multiple metric families, the model shows promise for biomedical text summarization.