TempLM: Distilling Language Models into Template-Based Generators

TL;DR

TempLM is a novel approach that distills pretrained language models into template-based generators, combining faithfulness and fluency, and significantly reducing unfaithfulness in out-of-domain text generation.

Contribution

The paper introduces TempLM, a method that effectively combines the strengths of PLMs and template systems through distillation, improving faithfulness and fluency in text generation.

Findings

TempLM outperforms original PLMs in faithfulness.

TempLM surpasses prior template systems in fluency.

TempLM reduces unfaithfulness rate from 83% to 0% on out-of-domain data.

Abstract

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.