Emotional analysis in texts is one of the difficult problems in text feature extraction. Semantic information is not unique for a large number of text features, which increases the difference in feature weight. Previous studies had dimensional disasters, loss of feature information, and a weak generalization ability in text feature extraction. To solve these problems, we first analyzed the advantages of support vector machines (SVMs) by multiple kernel learning (MKL). Then, an algorithm with local multiple kernel learning (LMKL) was proposed for a threshold model to select the locally optimal kernel function. It helped understand which text feature distinguishes emotions more effectively. Next, we analyzed the features of the local multiple kernel learning algorithm and discussed its generalization ability. The effectiveness of the method in this study was verified through comparison with other methods. The method reduced the feature dimensionality of the sample data set. Since the features with a weak classification ability were reduced, the accuracy of the classification was improved with increased efficiency.