Improving Compositional Generalization with Self-Training for Data-to-Text Generation
This addresses the costly data collection issue for data-to-text generation by improving few-shot generalization to novel compositions, though it is incremental as it builds on existing T5 models.
The paper tackles the problem of compositional generalization in few-shot data-to-text generation, showing that T5 models fail on unseen meaning representations and proposing a template-based input and self-training approach that improves tree accuracy by over 46% and reduces slot error rates by over 73% on benchmarks.
Data-to-text generation focuses on generating fluent natural language responses from structured meaning representations (MRs). Such representations are compositional and it is costly to collect responses for all possible combinations of atomic meaning schemata, thereby necessitating few-shot generalization to novel MRs. In this work, we systematically study the compositional generalization of the state-of-the-art T5 models in few-shot data-to-text tasks. We show that T5 models fail to generalize to unseen MRs, and we propose a template-based input representation that considerably improves the model's generalization capability. To further improve the model's performance, we propose an approach based on self-training using fine-tuned BLEURT for pseudo response selection. On the commonly-used SGD and Weather benchmarks, the proposed self-training approach improves tree accuracy by 46%+ and reduces the slot error rates by 73%+ over the strong T5 baselines in few-shot settings.