Template-Based Multi-solution Approach for Data-to-Text Generation

Data-to-text generation is usually defined into two parts: planning how to order and structure the information, and generating a text grammatically correct and fluent, that is faithful to the facts described in the input knowledge base source. A typically knowledge base consists of Resource Description Framework (RDF) triples which describe the entities and their relations. There are plenty of end-to-end solutions proposed to generate natural language descriptions from RDF, however, they require large and noise-free training datasets, lack control over how the text will be generated and there is no guarantee that the generated text verbalizes all and only the input. We address these problems by proposing a modular solution that uses templates and generates multiple texts over the data-to-text generation phases, returning the best one. Our experiments on a real-world dataset demonstrate that our approach generates higher quality texts and outperforms some baseline models regarding BLEU, METEOR, and TER.