Hierarchical Attention Networks for Document Classification

摘要: We propose a hierarchical attention network for document classification. Our model has two distinctive characteristics: (i) it structure that mirrors the of documents; (ii) levels mechanisms applied at wordand sentence-level, enabling to attend differentially more and less important content when constructing representation. Experiments conducted on six large scale text classification tasks demonstrate proposed architecture outperform previous methods by substantial margin. Visualization layers illustrates selects qualitatively informative words sentences.