融合BiLSTM的双图神经网络文本分类模型

采用图神经网络模型为整个语料库构建异构图处理文本分类任务时，存在难以泛化到新样本和词序信息缺失的问题。针对上述问题，提出了一种融合双图特征和上下文语义信息的文本分类模型。首先，为每个文本独立构建共现图和句法依存图，从而实现对新样本的归纳式学习，从双图角度捕获文本特征，解决忽略单词间依存关系的问题；其次，利用双向长短期记忆网络（bi-directional long short-term memory，BiLSTM）编码文本，解决忽略词序特征和难以捕捉上下文语义信息的问题；最后，融合双图特征，增强图神经网络模型的分类性能。在MR，Ohsumed，R8，R52数据集上的实验结果表明，相较于经典的文本分类模型，该模型能够提取更丰富的文本特征，在准确率上平均提高了2.17%，5.38%，0.61%，2.48%。

Dual graph neural networks with BiLSTM for text classification

When a heterogeneous graph was constructed for graph neural network model for the whole corpus to text classification, it was difficult to generalize to new samples and word order information was missed. To solve these problems, a text classification that incorporated dual graph features and contextual semantic information was proposed. Firstly, a syntactic dependency graph and a co-occurrence graph were constructed for individual text to inductive learning for new samples. Text features were captured from two perspectives, which solved the problem of ignoring dependency information between words. Secondly, bi-directional long short-term memory network (BiLSTM) was used to encode text, which solved the problem of ignoring word order features and having difficulty in capturing contextual semantic information. Finally, dual graph features were fused to enhance the classification performance of graph neural network model. The experimental results on MR, Ohsumed, R8 and R52 datasets show that the model can extract richer text features and have an average increase of 2.17%, 5.38%, 0.61% and 2.48% in accuracy compared with classical text classification models.