The main idea of this paper is to use the diffusion convolutional operator to learn text embedding that takes into account the global influence of the whole graph. It then incorporates the diffusion process in the loss function to capture high-order proximity. In contrast, previous works either neglect the semantic distance indicated from the graph, or fails to take into account the similarities of context influenced by global structural information. The author then conducts experiments on the task of multi-label classification of text and link prediction and shows that the proposed model outperforms the baselines. Strength: The high level idea of of this paper is good, and the method is novel.

It combines the advantages of a BERT model with a depends entirely on the syntactic dependency tree, but may not knowledge graph based synonym data. This synergy leverages work due to non-standard text.Tang et al. [9] divided sentences a dynamic attention mechanism to develop a knowledge-driven into left and right parts of perspective words and used two Long state vector. For classifying sentiments linked to specific Short Term Memory (LSTM) networks to model the correlation aspects, the approach constructs a memory bank integrating between perspective words and their left and right contexts, positional data. The data are then analyzed using a DCGRU to respectively. Huang et al. [10] used parameterized filters and pinpoint sentiment characteristics related to specific aspect threshold mechanisms to incorporate phase information into terms. Experiments on three widely used datasets demonstrate convolutional neural networks to effectively capture specific the superior performance of our method in sentiment classification.