Search and Learn: Improving Semantic Coverage for Data-to-Text Generation

Shailza Jolly, Andreas Dengel, Lili Mou

Abstract: Data-to-text generation aims to generate text description based on input data. Modern approaches are largely based on neural networks, which often require massive parallel data for training. Recently, researchers address few-shot data-to-text generation by fine-tuning pre-trained language models. In our work, we observe that such few-shot models suffer from the problem of low semantic coverage, i.e., important input slots are missing in the generated text. We therefore propose a search-and-learning approach that inserts the missing slots in a greedy manner and learns from the search results. Results show that our model achieves high performance on E2E and WikiBio datasets. Especially, we cover 98.5% of input slots on the E2E dataset, largely alleviating the low coverage problem.