基于语义感知的变长序列数据预处理框架

Deep learning frameworks generally adopt padding or truncation operations toward variable-length sequences in order to use efficient yet intensive batch training. However, padding leads to intensive memory consumption, and truncation inevitably loses the original semantic information. To address this dilemma, a variable-length sequence preprocessing framework based on semantic perception is proposed, which leverages a typical unsupervised learning method to reduce the different dimensionality to the exact size and minimize information loss. Under the theoretical umbrella of minimizing information loss, information entropy is adopted to measure the semantic richness, weights to variable-length representations is assigned, and the semantic richness is used to fuse them. Extensive experiments show that the information loss of the present strategy is less than the truncated embeddings, and the apparent superiority of the present method in gaining more information capability and achieving promising performance on several text classification datasets.