Meta Soft Prompting and Learning

Traditionally, either applying the hard prompt for sentences by handcrafting the prompt templates or directly optimizing the soft or continuous prompt may not sufficiently generalize for unseen domain data. This paper presents a parameter efficient learning for domain-agnostic soft prompt which is developed for few-shot unsupervised domain adaptation. A pre-trained language model (PLM) is frozen and utilized to extract knowledge for unseen domains in various language understanding tasks. The meta learning and optimization over a set of trainable soft tokens is performed by minimizing the cross-entropy loss for masked language model from support and query data in source and target domains, respectively, where the masked tokens for text category and random masking are predicted. The meta soft prompt is learned through a doubly-looped optimization for individual learners and a meta learner when implementing the unsupervised domain adaptation. The PLM is then closely adapted to compensate the domain shift in a target domain. The domain adaptation loss and the prompt-based classification loss are jointly minimized through meta learning. The experiments on multi-domain natural language understanding illustrate the merit of the proposed meta soft prompt in pre-trained language modeling under few-shot setting.