Large language models generate fluent texts and can follow natural languageinstructions to solve a wide range of tasks without task-specific training.Nevertheless, it is notoriously difficult to control their generation tosatisfy the various constraints required by different applications. In thiswork, we present InstructCTG, a controlled text generation framework thatincorporates different constraints by conditioning on natural languagedescriptions and demonstrations of the constraints. In particular, we firstextract the underlying constraints of natural texts through a combination ofoff-the-shelf NLP tools and simple heuristics. We then verbalize theconstraints into natural language instructions to form weakly supervisedtraining data. By prepending natural language descriptions of the constraintsand a few demonstrations, we fine-tune a pre-trained language model toincorporate various types of constraints. Compared to existing search-based orscore-based methods, InstructCTG is more flexible to different constraint typesand has a much smaller impact on the generation quality and speed because itdoes not modify the decoding procedure. Additionally, InstructCTG allows themodel to adapt to new constraints without re-training through the use offew-shot task generalization and in-context learning abilities ofinstruction-tuned language models.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)