A Comparative Study of Knowledge Graph-to-Text Generation Architectures in the Context of Conversational Agents

This work delves into the dynamic landscape of Knowledge Graph-to-text generation, where structured knowledge graphs are transformed into coherent natural language text. Three key architectural paradigms are explored: Graph Neural Networks (GNNs), Graph Transformers (GTs), and linearization with sequence-to-sequence models. We discuss the advantages and limitations of these architectures, and we do some experiments on these architectures. Performance evaluations on WebNLG V.2 demonstrate the superiority of sequence-to-sequence Transformer-based models, especially when enriched with structural information from the graph. Despite being unsupervised, the CycleGT model also outperforms GNNs and GTs. However, practical constraints, such as computational efficiency and model validity, make sequence-to-sequence models the preferred choice for real-time conversational agents. Future research directions include enhancing the efficiency of GNNs and GTs, addressing scalability issues, handling multimodal knowledge graphs, improving interpretability, and devising data labeling strategies for domain-specific models. Cross-lingual and multilingual extensions can further broaden the applicability of these models in diverse linguistic contexts. In conclusion, the choice of architecture should align with specific task requirements and application constraints, and the field offers promising prospects for continued innovation and refinement.