Research On Technologies Of Special Reports Auto-Summarization

How to compress, integrate and extract the core content of massive information has become a hot research area in recent years. Composing special reports is an important approach to achieve this goal. Although the technologies of special reports automatic generation have been extensively studied, there is still a large gap between automatically generated and human-written special reports on readability and accuracy, which indicated a great room for improvement. Therefore, we carried out researches on three key technologies of special reports automatic generation in this paper.(1) First, we studied image feature expression method based on Dictionary Learning and Min-Hash approaches. We obtained dimension-reduced representations of original image features by exploiting these techniques, which provided basic materials for special reports’ data clustering.(2) After achieving feature expressions of text and images, we proposed a semi-supervised hierarchical topic model in order to clustering text and image by topics. We use topic modeling techniques to cluster images and text by topics, which provided the materials for topic content analysis and synthesis.(3) After achieving topic-related document sets, we proposed an automatic topic summarization approach based on deep learning approaches. We used Deep Learning techniques for topic-related documents analysis and synthesis and finally generated summaries automatically which could cover key content in topics.We carried out serveral experiments for the three methods described above on multiple datasets. Our methods have shown better performances than the control methods on most of the evaluation metrics, which indicates that our methods could well satisfy the application requirements of special reports automatic generation.