Research On Text-to-Scene Conversion Technology Based On Entity And Spatial Relationship Reasoning

The concept of "Metaverse" has been constantly mentioned and widely applied in various fields such as education,medical treatment,real estate,etc.since its inception,and virtual scene construction is the foundation for building the "Metaverse".Faced with the increasing demand for virtual scenes,traditional virtual scene construction requires professionals to create and the process is complex and time-consuming,and there is an urgent need for new technologies to change the current situation.With the continuous development and integration of natural language processing and virtual reality technology,"text-to-scene conversion" technology as a new field is constantly being studied.It can visualize the scenes described by natural language,so that non-professionals can quickly build virtual scenes.At present,the research and prototype system of Chinese "text-to-scene conversion" technology is still in its infancy.The generated scenes are single,and the model is placed in disorder,which is far from user expectations.Aiming at the above problems,the research on text-to-scene conversion technology based on entity and spatial relationship reasoning is proposed.Finally,a prototype system for text-to-scene conversion of the Riverside Scene at Qingming Festival is built according to the research results to verify the feasibility of the method and explore the application of "Metaverse" in the field of cultural protection.The following are the main research contents:Firstly,for the generated single scene problem,this paper uses Hidden Markov Model(HMM)to deduce hidden entity.Collect corpus of the text describing the scene of Riverside Scene at Qingming Festival,create the orientation dictionary,use the HMM to further reason the scene entity,dig out the words associated with the scene entity,judge the relevance between the scene entity and related words by calculating the expected value,and control the prediction degree of the scene by adjusting the threshold value in the scene generation.Finally,the whole process of extracting visual scene entity and spatial relation from text is elaborated systematically.Secondly,to solve the problem of chaotic placement of the generated scene model,this paper proposes an algorithm based on the prior probability scene layout and fuzzy distance calculation between the models,so as to supplement the missing spatial relationship information.Create a spatial relation library,define 8 basic spatial constraints,introduce the specific application of two algorithms in scene generation in detail,solve the problem of how to place the model when the spatial relation information is missing in the scene description text,and systematically describe the construction process of text-to-scene conversion.Finally,a scene generation system of Riverside Scene at Qingming Festival is designed according to the above methods,and the process from text input to scene generation of Riverside scene at Qingming Festival is completed.Design and develop the system,create a model library and preprocess the model,conduct semantic analysis of the input scene description text using word segmentation tools,extract the scene entity information and spatial relationship triplet,get the words associated with the scene entity,and obtain the corresponding model range according to the expected size to enrich the scene content.After that,the model matching algorithm is used to obtain the model that is similar to the preset generated entity in the model library to realize the placement of scene objects,and finally make the generated scene more abundant and reasonable layout.