TempLM: Distilling Language Models into Template-Based Generators
This addresses the issue of generating faithful and fluent text for data-to-text tasks, though it is incremental as it combines existing distillation and template-based approaches.
The authors tackled the problem of unfaithful or inappropriate content generated by pretrained language models by distilling a PLM into a template-based generator, resulting in TempLM reducing unfaithfulness from 83% to 0% on an out-of-domain evaluation and improving fluency over prior template systems.
While pretrained language models (PLMs) have greatly improved text generation, they have also been known to produce unfaithful or inappropriate content. In contrast, classic template-based systems provide strong guarantees of faithfulness at the cost of fluency. We propose TempLM, which achieves the best of both worlds by distilling a PLM into a template-based generator. On the E2E and SynthBio data-to-text datasets, we show that TempLM is more faithful than the original PLM and is more fluent than prior template systems. Notably, on an out-of-domain evaluation, TempLM reduces a finetuned BART model's unfaithfulness rate from 83% to 0%. In a human study, we find that TempLM's templates substantially improve upon human-written ones in BERTScore.