| Text segmentation is to obtain the boundaries between segmentations vvhieh meet the requirements of larger semantic consistency in the segmentation and smaller semantic consistency between segmentations according to topic similarity. This paper presents an improved text segmentation model, which uses the topic model as a similarity measurement method, and uses the improved graphcut algorithm to search the segmentation boundary.After overviewing the previous semantic similarity measure, we first discuss the LDA, HDP&LDA and CTM, and then try to use the three topic models to semantic similarity measure. Obtaining of topics is the basis of text segmentation. Since vocabulary is the basic semantic units of text, shallow semantic information can be got by using the simple statistics on the vocabulary of the text, and deep semantic information can be got by using the topic model. The research shows that using of LDA in text segmentation can improve the results. Compared to LDA, HDP can determine the property topics number: CTM can represent correlations among topics. Through the contrast experiments, we verify that setting the property number of topics in the topic model and using the correlations of topics can improve the effect of text segmentation.In this paper, we use graphcut algorithm which is widely used in Computer vision as text segmentation boundary search method. There is a large amount of text segmentation methods which need artificial priori parameters. These parameters tend to have a significant impact segmentation results and cannot be given desired value,(irapheut algorithm can effectively avoid this problem. In order to make the graphcut algorithm suitable to text segmentation task, this paper has done the following three aspects: lust, redefine the text segmentation border search problem as the expression of energy function that can be solved by graphcut algorithm; then redefine the graph cut algorithm dynamical graph-building step, according to the linear nature ol’ the problem; finally, the feasibility analysis. Finally, we discuss details of the improved text segmentation model based on the topic models and graphcut algorithm and it’s implementation. And then, according to the analysis of the graphcut algorithm, we combine the idea of simulated annealing with the graphcut algorithm to improve the graphcut algorithm. Kxperimental results demonstrate that the models have improved performance on text segmentation task. |
