Conference summary generation method based on LDA and Transformer

Automatic text summarization technology has shown extensive potential across various domains. Despite significant advancements by deep learning in improving text summaries, several issues persist. In extraction methods, effective modeling of selection and ordering of extraction units is challenging. Generative methods face problems like inadequate key information modeling, excessive out-of-set words, numerous model parameters, and slow processing speeds. To address these, this paper proposes the BTN method, which combines extraction and generative approaches. It uses the LDA model and TF-IDF text characteristics, followed by a neural network to capture sentence relationships, generating summaries with reduced redundancy and increased diversity. Additionally, the accuracy of graph ranking real-time summarization is enhanced, and SBERT is incorporated to improve the extractive summarization. To further enhance accuracy and diversity, BART is integrated into the Seq2Seq generation model, constructing a abstractive summarization. Experimental results on the CNN/Daily Mail dataset show that the proposed algorithm significantly outperforms traditional methods.