Attention-Based Bidirectional Hierarchical LSTM Networks for Text Semantic Classification
Xiaoming Shi, Ran Lu · 2019
Recently, in the semantic classification task at the document level, the traditional semantic classification method still has challenges in coding the hierarchical relationship between word level and sentence level. Aiming at this problem, we construct a bidirectional hierarchical LSTM network model (HBLSTM-ATT) based on the attention mechanism, and calculate the correlation before and after the sentence. At the same time, this paper focuses on document-level classification, using hierarchical structures to build document-level vectors from word vectors. Experimental results on Yelp, IMDB, Yahoo Answer and Amazon review datasets show that the proposed method is efficient in text semantic classification tasks compared to previous machine learning methods, and the semantic classification methods for gated neural networks. The accuracy rate for classification and stability have robust superiority over competitors and achieve the state-of-the-art results.