In-Context Learning是最近比较火热的方向,其主要针对超大规模模型(例如1750B参数量的GPT-3模型),在只提供少量标注样本作为提示的前提下,即可以实现很惊艳的效果。本文将元学习引入到In-Context Learning中。
论文PDF:https://doi.org/10.18653/v1/2022.acl-long.53
本文提出in-context tuning(ICT)用于few-shot learning
For example, to coax the model into performing sentiment classification on the target input “This movie is a waste of time”, we prompt the LM with the sequence “I like the movie! Positive review? Yes. Horrible Movie! Positive review? No. This movie is a waste of time. Positive review? ___”, and predict “positive” if the next word is more likely to be “Yes” rather than “No”.
raw LMs are not optimized for in-context FSL during pre-training, and exhibit undesirable behavior when used for FSL
Previous work has also shown that prompting raw LMs is often oversensitive to example choices and instruction wording
在训练(fine-tuning)阶段,给定一系列的训练task,每一个task都有相应的instruction,以及该task对应的少量样本(输入/输出对)。在测试阶段,给定一个新的unseen task,以及该task对应的instruction和少量样本(输入/输出对),旨在让模型能够对测试样本预测其类别。如下图,给定一个情感分析task:
在训练时,直接对instruction ITI_TIT、若干少量标注数据 STS_TST 以及target样本 xtargetTx_{target}^TxtargetT 进行拼接,并基于in-context learning训练目标进行优化,预测对应类别 ytargetTy_{target}^TytargetT:
(1)baseline模型
Raw In-context Learning
与in-context tuning一样,给定unseen task的instruction、少量标注样本(输入/输出对)和样本输入,直接预测样本的输出。此时没有fine-tuning过程,属于zero-shot learning场景。
Instruction-tuning + Fine-tuning
给定若干种类的task,基于instruction以及样本的输入,用于训练。在fine-tuning阶段,给定unseen task的instruction以及K个样本,进行fine-tuning。
Instruction-tuning
此时只有instruction和样本的输入,属于zero-shot learning场景。
MAML
给定若干task的instruction和一个样本输入,用于训练并预测目标。其与传统的MAML一样,只是训练目标变为instruction tuning