| 摘 要: 针对传统词向量模型无法获取完整的语义表达，以及基础神经网络模型未能兼顾提取多种关联特征等问题，提出了一种融合预训练语言模型(ERNIE)和深层金字塔神经网络结构(DPCNN)/双向门控循环单元-注意力机制(BiGRU-Attention)的双通道文本情感分类模型。基于DPCNN的左通道负责提取文本长距离依赖表示，基于BiGRUAttention的右通道负责提取文本时间序列特征和关键信息。此外，均使用ERNIE模型提供动态字向量。最后，拼接融合双通道中的信息特征以获取最终的文本表示。实验结果表明，ERNIE-DBGA模型的准确率最高达到97.05%，优于其他对比方法，验证该模型可以有效提升情感分类的性能。 |
| 关键词: 文本情感分类；ERNIE；双通道；DPCNN；BiGRU；注意力机制 |
|
中图分类号: TP391
文献标识码: A
|
|
| Text Sentiment Classification Model based on ERNIE and Fusion of Dual-Channel Features |
|
YAO Huanhuan, ZHU Xiaodong
|
(Business School, University of Shanghai for Science and Technology, Shanghai 200093, China)
yaohhh1011@163.com; zhuxd@usst.edu.cn
|
| Abstract: In response to the problem that the traditional word vector model cannot obtain the complete semantic expression, and the basic neural network model cannot take into account the extraction of multiple associated features, this paper proposes a dual-channel sentiment classification model combining ERNIE (a pre-training language representation model enhanced by knowledge) and DPCNN (Deep Pyramid Convolutional Neural Networks) / BiGRU (Bidirectional Gated Recurrent Unit)-Attention (ERNIE-DBGA). The left channel based on DPCNN is responsible for extracting the long-distance dependency representation of the text, while the right channel based on BiGRU-Attention is responsible for extracting the text time series features and key information. In addition, the ERNIE model is used to provide dynamic word vectors. Finally, the information features in the two channels are concatenated and fused to obtain the final text representation. The experimental results show that the accuracy of ERNIE-DBGA model is up to 97.05%, which is superior to other comparison methods, verifying that the model can effectively improve the performance of sentiment classification. |
| Keywords: text sentiment classification; ERNIE; dual-channel; DPCNN; BiGRU; attention mechanism |
