Currently, an enormous amount of textual data is generated in the biomedical domain due to the rapid spread of the Internet and the advancement in information technology. Users and researchers often have difficulty finding valuable information among the vast amount of data. It takes longer to search large databases of unrelated and related information, and data loss also occurs during this process. Text summarization systems are used to create information summaries automatically. A subset of the original text is produced with relevant information and topics without redundancy in automatic text summarization. Our study proposes a multilayer graph-based biomedical text summarization. A comprehensive graph was built from the given texts using semantic concepts and a multilayer graph. Tools such as MetaMap and OGER allow the extraction of concepts from the UMLS database, and SemRep can identify relationships between concepts. The result is three-layered: 1) Similarity-based on words, 2) Similarity-based on semantic concepts, and 3) Similarity-based on the co-occurrence of words. There have been various methods proposed for domain-independent summarization and biomedical texts. General summarization methods do not work well in the biomedical domain. Compared to other texts, biomedical texts contain many complex concepts, synonyms, abbreviations. Algorithms such as machine learning, graph modeling, optimization, and statistical methods have been used in biomedical text summarization. Different methods have demonstrated that domain-specific knowledge and semantic concepts produce more accurate summaries. The previous graph-based methods presented relationships in the text with simple graphs and one-dimensional views. As a starting point, it is important to create a richer graph that includes as many different and effective relationships as possible. Based on multilayer graphs, several methods for clustering and scoring sentences have been investigated. As a result, methods were chosen to produce a richer output while maintaining the relationships and information gleaned from the text. The approach used at this stage is clustering and ranking. Clustering refers to the creation of different groups of sentences based on multilayer graph clustering. A variety of heuristics are then used to extract the output sentences. As a part of defining heuristics, it is essential to understand how statistical features can be used for selection. In the sentence ranking process, multi-layered graphs are produced without clustering and with PageRank algorithms. Using this approach, we present two versions. Based on the ranking algorithm output list, the first version includes the top sentences. The second version involves calculating sentence length based on the number of concepts. We add this parameter as an effective parameter in our proposed method for selecting sentences. eva‎luation of the proposed method has been conducted with common approaches and tools in this field. The ROUGE and BertScore standard criteria, which compare and score the generated summary text to the article abstract as a golden summary, are two popular approaches. In the study, different algorithms and heuristics are eva‎luated, and the level of improvement is reported

ناصر قديري مدرس

مريم لطفي شهرضا، فابيو رينالدي