Feature Extension for Chinese Short Text Classification Based on Topical N-Grams

Because of the feature sparseness problem, conventional text classification methods hardly achieve a good effect on short texts. This paper presents a novel feature extension method based on the TNG model to solve this problem. This algorithm can infers not only the unigram words distribution but also the phrases distribution on each topic. We can build a feature extension library using TNG algorithm. Base on the original features in short texts, we can compute the topic tendency for each of these texts. According to the topic tendency, the appropriate candidate words and phrases are selected from the feature extension library. And then these candidate words and phrases are put into original short texts. After extending features, we use the LDA and SVM algorithm to classify these expanded short texts and use precision, recall and F1-score to evaluate the effect of classification. The result shows that our method can significantly improve classification performance.