Obviously, compared to three feature sets experimental results, the precision and recall of deception detection changes significantly, with the precision from 0.79 to 0.85, and the recall from 0.75 to 0.84. Simultaneity, the result shows the precision is the highest as using the second feature set. In the experiments the other important problem found is that the proportion and scale of training corpus and test corpus have an effect on the experimental results, so in order to test the effect we choose a different proportion of training corpus and test corpus for experiments, including 2 to 8, 4 to 6, 5 to 5, 6 to 4 and 8 to 2 of the proportion of training corpus and test corpus. To five types corpus we all conduct the homologous experiments by using SVM classifier and the same feature set including the words whose CHI statistics feature value is bigger than 3 and the linguistics cues whose feature value is bigger than 0.3. Fig.3 shows the trends of the recall and precision of the experiments using five types corpus.