The Research Of Scene Related Entity Reasoning Based On Word Frequency

The text-to-scene conversion is a process that is dependent on computer to realize the creation of 3D scenes from natural language description automatically.It has a broad application prospect in the fields of artificial intelligence,animation,education,military and so on.The scene generated by the existing text-to-scene conversion system is too simple and contains only the entities described in the text.The implicit elements in the scene are insufficiently expressed and the generated scene is not realistic enough.Firstly,based on the existing researches on text-to-scene conversion and natural language processing,this thesis expounds the definition of scene and gives the concepts of scene entity model and scene entity dictionary,which lays a foundation for further research.Secondly,aiming at the problem of inadequate representation of implicit elements in scene,this thesis proposes a method of scene associative entity reasoning based on word frequency.Based on the existing natural language processing techniques,the segmentation and part-of-speech tagging,phrasing and semantic analyzing and so on are carried out in this thesis to realize the recognition of the scene entity word information.Based on a large number of document data on the Internet,this thesis uses word frequency analysis to obtain the scene associated entity word sequence set,and combines the inverse document frequency to optimize the entity word sequence set,and verifies the effectiveness of the method through experiments.Then,based on the co-occurrence frequency analysis of the words,we find that there is a mixed problem of scene classification,and the problem can not be solved effectively from the co-occurrence frequency of words.Then,combining with statistical data,the reasoning method of the optimal scene related entity set is proposed.First,the problems of using the scene entity sequence set directly as the result of the scene related entity reasoning are analyzed and this thesis proposes a scene classification analysis method combining LDA and optimal topic number.Secondly,this thesis filters the low TF-SIDF weight entity words by setting the weight threshold.In this way,we realized the construction of the scene core entity word set,the scene expansion entity word set and the optimized scene entity word set,and set up the scene entity dictionary by using the best scene entity word set.Finally,an experimental prototype system is established.On the basis of building a visual entity name library and a model library,this thesis uses Java 3D graphics technology,sets up the experimental environment,and introduces the scene entity dictionary library.In contrast to the existing text conversion system results,this thesis proves the feasibility of method and technical route of this thesis proposed and effectiveness.