A divide-and-conquer approach to neural natural language generation from structured data

Current approaches that generate text from linked data for complex real-world domains can face problems including rich and sparse vocabularies as well learning examples of long varied sequences. In this article, we propose a novel divide-and-conquer approach automatically induces hierarchy “generation spaces” dataset semantic concepts texts. Generation spaces are based on notion similarity partial knowledge graphs represent the domain feed into sequence-to-sequence or memory-to-sequence learners concept-to-text generation. An advantage our is models exposed to most relevant during training which avoid bias towards majority samples. We evaluate two common benchmark datasets compare hierarchical against flat setup. also conduct comparison between models. Experiments show overcomes issues sparsity learns robust lexico-syntactic patterns, consistently outperforming baselines previous work by up 30%. find while outperform in some cases, latter generally more stable their performance safer overall choice.