Research And Analysis On Scene Identification From Text

In Text–to-Scene Conversion, to draw the 3D renderings from the plot in the text, we must make use of natural language processing technology, and comprehensively understand the information in the scene of the text. Therefore, we propose to do a research on scene identification for the "Text-to-Scene Conversion".The main work of this paper is as follows:1. Definition of terminology and construction of the scene categories. According to the research needs, we define the terminology for the scene idendificaiton. We propose an algorithm which is for clustering the scenes information words from the childen stories based on the sememe hypernym and concept similarity of HowNet, and gets 48 scene categories architecture with manual adjustment. We regard it as the classes for the scene identification.2. The construction of the knowledge base. We preprocess the corpus, introduce the Relevant Measurements formulas, and give the formal definition. We present relevant measurements to identify scenes in natural language texts with the action information and the scene categories. We use four kinds of measurement MI, Cosine correlation coefficient,χ2 test, and Dunning's likelihood, to generate the knowledge base. The knowledge base consists of some records, such as (action, scene category, values). The values are sorted by descending. Finally, we construct the knowledge base for the time categories and the season categories similarly. The knowledge base is regarded as the basis for the scene identification.3. Scene identification. This paper uses 1-best to identify the scenes in the text. 1-best emphasizes the impact of a single record on the knowledge base, and expands the interference of noise on the identification. We propose a feasible method―vote‖to solve the problem. The results of the testset identified by the relevant measurements are significant on the t-test. We make use of precision, recall, F score to evaluate the inferring results on the scene categories for the first time, and the experimental results indicate that corpus expansion based on common sense is feasible. We analyze the scene identification by F-Score quantitatively on multiple contexts, knowledge bases and identification methods, and it indicates that the descriptions of adjacent sentences on the same scene are correlative and Dunning's likelihood is better than other methods for scene identification.4. System design and implementation. We implement a system for scene identification, which is based on statistical method to infer the locations, times and seasons for the Chinese factual text. Users can participate in interactive events, and improve the results of the identification. The system has a significant effect.