Abstract: Temporal relation recognition is a research focus in NLP(nature language processing). This paper identifies temporal relations based on temporal segment and topic segment, which semantic granularities were coarser. First, temporal segments were recognized according to temporal discourse characters. Then, topic segments were recognized through computing similarity between paragraphs and the SVM model. Final, within each topic segment, temporal relations between the adjacent temporal segments were identified by maximum entropy classifier. Experiments were made on TempEval-2010 corpus of Chinese, the macro-average precision of temporal relation recognition was 60.09%. The experimental results show that introduction of temporal segments can reduce the redundant recognition of the temporal relations between events. And with the scope constraint of topic segments, the results of temporal relations become more concise and understandable.