Semantic Matching-Based Neural Network for Text Summarization in Knowledge Extraction

Text summarization techniques represent an essential research frontier in natural language processing, focusing on distilling lengthy texts into concise summaries that retain key insights. Traditional extractive summarization methods primarily operate at the level of isolated sentences and are limited in modeling the semantic relationships between a document and its summary, which often leads to redundant, fragmented and weakly informative outputs in knowledge intensive settings. To address these gaps, this paper proposes Candidates Matching Text Summarization, a semantic matching-based neural network for extractive text summarization in knowledge extraction scenarios. This method constructs candidate summary sets by coherently combining multiple pivotal statements from original documents, thus preserving the critical contextual connections imperative. Subsequently, the method captures linguistic and topical representations of texts using a fine-tuned BERT model and a neural topic model, respectively. A semantic matching network then evaluates the semantic congruence between each candidate summary and the full document, selecting summaries that are most aligned with the original content's significance. Experimental validations conducted on relevant datasets demonstrate that this approach achieves up to a 6.47% performance enhancement compared to conventional summarization models. This improvement demonstrates that the method achieves more accurate semantic understanding and higher-quality summary generation, which can be practically applied to enhance knowledge extraction, organizational information processing and knowledge sharing efficiency.