Related Work on Few-shot Method: A Review

Abstract

Abstract. As Natural Language Processing (NLP) technologies continue to expand into novel domains, few-shot learning emerges as a pivotal approach to addressing the challenge of data scarcity. Traditional neural networks, due to their strong reliance on abundant data, have posed limitations in their applicability to new domains. Consequently, there is a pressing need to introduce fresh research perspectives and solutions within this realm, aiming to propel its development towards greater practicality and efficiency. Firstly, solving the zero or few-shot Dialogue State Tracking problem has become necessary as the demand for deploying such systems in new domains continues to grow. This article explores the performance of D-REPTILE, a meta learner for Destination Address (DST) problems, in unknown domains. Then PET is extensively studied. Real-world tests on RAFT show cue-based learning works in low-sample settings, reinforcing the importance of instruction-based learning for human-like few-shot capabilities. Thirdly, to address the overfitting problem, this paper also explores the LA- UCL model, as well as its application, development, and challenges, which enables the LA-UCL model to enhance the Large Language Model data expansion effect through two modules. Finally, CausalCollab is introduced, which uses Incremental Stylistic Effects (ISE) as a guiding estimator for assessing the effectiveness of LM-human cooperation tactics through time.