Does the Order of Training Samples Matter? Improving Neural Data-to-Text Generation with Curriculum Learning
This work provides a strong specific gain in training efficiency and performance for researchers and practitioners working on neural data-to-text generation.
This paper investigates curriculum learning for neural data-to-text generation, where training samples are ordered based on the model's competence. They propose a soft edit distance metric for ranking samples, achieving a 38.7% reduction in training time and a 4.84 BLEU score boost.
Recent advancements in data-to-text generation largely take on the form of neural end-to-end systems. Efforts have been dedicated to improving text generation systems by changing the order of training samples in a process known as curriculum learning. Past research on sequence-to-sequence learning showed that curriculum learning helps to improve both the performance and convergence speed. In this work, we delve into the same idea surrounding the training samples consisting of structured data and text pairs, where at each update, the curriculum framework selects training samples based on the model's competence. Specifically, we experiment with various difficulty metrics and put forward a soft edit distance metric for ranking training samples. Our benchmarks show faster convergence speed where training time is reduced by 38.7% and performance is boosted by 4.84 BLEU.