Bidirectional long short-term memory(BiLSTM) and its variants can handle variable length sequences. Due to the complex semantic information of text and the high dimension of embedded text data, BiLSTM shows a weak ability of low-level network learning. And the problem of network degradation is easy to occur when learning high-level feature representation using overlay network layer. In order to solve these problems, this paper proposed a closed-loop BiLSTM module to enrich the semantic representation of hidden state in each layer of network structure. Meanwhile, it used residual connection and enhanced sparse representation strategy to optimize the module, and sparsed the feature vectors of hidden state to alleviate network degradation. Finally, this paper adopted the weighted fusion of multi-channel word embedding and realized the fusion of semantic information and emotional information under the low-dimensional tensor to enrich the text representation of the input layer. This paper verified the datasets of emotional classification and problem classification by experiments. The results show that the proposed model has excellent performance, good classification accuracy and robustness in capturing the expression of text's emotion information.

[2021-06-05] Printed Article /doc/2020.06.0161/v2