| With the development of Internet technology,the explosive growth of network data has made it difficult for people to efficiently obtain information resources,leading to increasing attention being paid to automatic text summarization technology.Unlike general text summarization,conference text contains multiple participants and longer content.Since the outbreak of the COVID-19 pandemic,the demand for online conferences has sharply increased,making automatic summarization of meetings even more important.Compared to extractive summarization methods,generative summarization methods are more conducive to people’s understanding and summarization of conference content,resulting in better summarization effects and more suitable results.Therefore,summarizing meetings through automatic text generation can provide high-quality and convenient services for conference participants and recorders,helping people better obtain the main content of meetings.For research on Chinese conference summary generation models,a Chinese conference summary generation model based on dialogue behavior is proposed,and a research dataset is constructed using a combination of machine summarization and manual screening.By analyzing dialogue semantics,a bidirectional long short-term memory network(Bi-LSTM)and attention mechanism are proposed to construct a dialogue behavior labeler,used to obtain the dialogue behavior of each sentence in the conference,to explore participants’ interaction intentions,and then use a double-layer transformer structure for Chinese conference summary generation.Finally,the effectiveness of the improvement is demonstrated through comparative experiments on the dataset.For the optimization of conference summary models combining semantic and structural information,a semantic and structural information combined conference summary optimization model based on knowledge graphs is proposed.Two layers of knowledge graph structures are designed to obtain the conference’s structural information,and a hierarchical conference summarization model based on dialogue behavior is used to learn the conference’s semantic information.The structural information and semantic information are combined as the final conference representation and input to the decoder component to generate the conference summary,achieving the optimization of the conference summary model that combines semantic and structural information.Finally,from the perspective of software engineering,an intelligent conference summary generation prototype system based on the SpringBoot development framework is designed and built.The conference summary generation function is implemented,And through functional testing,robustness testing and performance testing to verify that the system has achieved the expected results,helping users more conveniently and quickly use the intelligent conference summary generation system. |
