Multistage Collaborative Knowledge Distillation from a Large Language Model for Semi-Supervised Sequence Generation

We study semi-supervised sequence generation tasks, where the few labeledexamples are too scarce to finetune a model, and meanwhile, few-shot promptedlarge language models (LLMs) exhibit room for improvement. In this paper, wepresent the discovery that a student model distilled from a few-shot promptedLLM can commonly generalize better than its teacher to unseen examples on suchtasks. We find that the student is able to learn a general pattern from thehigh-quality pseudolabels produced by the teacher during knowledge distillation(KD), and favorably not a general pattern from the low-quality pseudolables.Leveraging this discovery, we propose a new method, Multistage CollaborativeKnowledge Distillation from an LLM (MCKD), for these tasks. MCKD first few-shotprompts an LLM to produce pseudolabels for unlabeled data. Then at each stageof an iterative KD process, a new pair of students is trained on disjointpartitions of the pseudolabeled data, and produces new and improvedpseudolabels for their unseen partitions. We conduct extensive experiments onfour syntactic and semantic parsing datasets and show the effectiveness of MCKDfor low-resource semi-supervised sequence generation. On CRAFT biomedicalparsing, for example, 3-stage MCKD with 50 labeled examples outperforms an LLMteacher and vanilla KD by 7.5the performance of supervised finetuning with 500 labeled examples.