Chinese Short Text Classification Based on Interactive Attention Mechanism

With the explosive growth of social media, millions of Chinese short texts are produced every day. However, these short texts is exactly short, so the machine can not extract features accurately from short texts for classification learning. In this paper, we propose an effective semantic enhancement model that not only combines character-level features, word-level features and sentence-level features as additional features for Chinese short text, but also removes noise by introducing Interactive Attention Mechanism to measure the semantic similarity between the i-th character and short text representation. Experiments on the open dataset Chinese news headlines classification task (190,000 news headlines/18 tags) show that our model achieves 84.53 %, 84.32 % and 84.36 % in macro-averaged precision, recall, F1, which boosts 1.26 % F1, 1.22 % recall and 1.3 % precision and achieves the state-of-the art performance.