A novel topic segmentation algorithm is proposed in this paper.This algorithm first partitions text into some blocks.After that it constructs whole-length lexical chains to analyze multiple subtopics of this text.By constructing graph which describes blocks covering subtopics

the similar blocks which describe same subtopic can be classified.In order to solve the situations that segmentation points drop inside blocks

it segments blocks again.Experiment results demonstrate that by analyzing topic of text

this algorithm can remove interferences

which are aroused by irrelative features

from segmentation results.By constructing graph which describes blocks covering subtopics

it can mix similarities of adjacent and disconnected blocks together

and increases segmentation precision.The second segmentation makes segmentation results more reasonable.