Contrastive Learning Fusing Dual Coding for Knowledge Graph-to-Text Generation

The knowledge graph-to-text (KG-to-text) generation task aims to generate the corresponding high-quality text based on a given knowledge graph. Most existing methods linearize the graph structure to assist in fine-tuning pre-trained language models for KG-to-text generation. However, the simple linearization is missing the rich structural information of the graph, and there is no effective strategy to fuse the graph structure and contextual information. To address the above problem, we propose for the first time the fusion of structural information of graphs with textual information based on a comparative learning framework. The approach includes a coding module, a comparative learning module and a decoding module. In this case, the encoding module captures the structural and textual information of the graph to obtain an output of two encodings, respectively. The comparative learning module fuses the two encodings by learning from positive and negative samples. Finally, the decoding module is introduced to achieve decoding while constraining the model from generating irrelevant text. Experimental results on two benchmark datasets show that our method obtains competitive performance.