Abstract: To capture the semantic inference result between different text fragments and resolve the reasoning problem of the text fragments,the classifier method was adopted to implement the entailment decision and the feature selection was the important factor influencing the classifier performance.The lexical features were applied in the system as the baseline and then the syntactic feature and the semantic feature were joined.The construction of lexical chains could mine the semantic relation between T and H and it had been used on the classifier to verify its effectiveness.The system performance had been evaluated on the data set of RTE-3 ~ RTE-5.The classifier of AdaBoost and SVM achieved the higher precision of 61.0% and 61.8%,respectively.The t-test results indicate that the semantic feature based on the lexical chain makes the system performance improve significantly.