Tagging And Retrieving Documentaries Based On Deep Neural Network

Scientific documentaries are always welcomed in elementary or secondary schools.It can not only help teaching,but also impress students.However,it is difficult for teachers to find other high-definition video material besides the unified compact discs companied with teaching reference books.Teaching video material has two requirements:(1)short,(about 3-5 minutes);(2)fit to the textbook.Scientific documentaries are a neglected library of high-definition video.A scientific documentary is ignored because it always takes too long(about 45-90 minutes),and has own agenda.Therefore,we propose a prototype system called ASTROS,which can help teachers find short materials for teaching by scientific documentaries.ASTROS contains three main components:knowledge map extraction,documentary subtitle tagging,and hit re-ranking.The knowledge map extraction module is used to extract knowledge map from textbooks and assign different weights to different levels of concepts.By using LSTM,the documentary subtitle-tagging module will tag subtitles in documentaries and associate them with knowledge graphs.The hit re-ranking module will re-rank multiple hits in the search according to different demands.We have three innovations:(1)Classifying subtitles by using deep neural networks,and associating subtitle sentences with knowledge map to achieve the semantic association between documentaries and textbooks,the accuracy of neural network classification reaches 96%.(2)Improving searching algorithm,adding the upper-word association rules on the principle of finding participle matching items,so that the semantically related search terms can appear in the search results and it is more suitable for educational short text search.The average value reaches 0.89.(3)Splitting the documentary by subtitles to facilitate teachers to obtain examples in the documentary.In summary,we builds a knowledge map based on the content of the textbook,and uses the nodes in the knowledge map as labels to establish a deep neural network to classify the subtitle sentences.Semantic related items are possible to appear in the search results through the improved reordering algorithm.Preliminary results show that it can facilitate teachers to prepare their courses.