A semantic-extension-based algorithm for short texts is proposed, by involving the Word2vec and the LDA model, to improve the performance of classification, which is frequently deteriorated by semantic dependencies and scarcity of features. For every keyword within a short text, weighted synonyms and related words can be generated by the Word2Vec and LDA model, respectively, and subsequently be inserted to extend the short text to a reasonable length. We not only have established a criterion by means of similarity estimation to determine whether a sentence should be extended, we designed a scheme to choose the number of extended words. The extended text will be classified. Experimental results show that, the classification performance of the proposed algorithm, in terms of the precision rate, is approximately 5％ higher than that of the TF-IDF model and approximately 10％higher than that of the VSM method.