Robustness of Demonstration-based Learning Under Limited Data Scenario

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1769 - 1782 December 7-11, 2022 ©2022 Association for Computational Linguistics Robustness of Demonstration-based Learning Under Limited Data Scenario Hongxin Zhang1, Yanzhe Zhang2, Ruiyi Zhang3, Diyi Yang4 1Shanghai Jiao Tong University, 2Georgia Institute of Technology 3Adobe Research, 4Stanford University 1icefox@sjtu.edu.cn, 2z_yanzhe@gatech.edu 3ruizhang@adobe.com, 4diyiy@cs.stanford.edu Abstract Demonstration-based learning has shown great potential in stimulating pretrained language models’ ability under limited data scenario. Simply augmenting the input with some demon- strations can significantly improve performance on few-shot NER. However, why such demon- strations are beneficial remains unclear since there is no explicit alignment between the demonstrations and the predictions. In this pa- per, we design pathological demonstrations by gradually removing intuitively useful informa- tion from the standard ones to take a deep dive of the robustness of demonstration-based se- quence labeling and show that (1) demonstra- tions composed of random tokens still make the model a better few-shot learner; (2) the length of random demonstrations and the relevance of random tokens are the main factors affecting the performance; (3) demonstrations increase the confidence of model predictions on cap- tured superficial patterns. We have publicly released our code at https://github.com/ SALT-NLP/RobustDemo. 1 Introduction Current large pretrained language models (PLMs) struggle to learn NLP tasks under limited data sce- narios (Devlin et al., 2019; Liu et al., 2019; Lewis et al., 2020; Xie et al., 2020; Huang et al., 2021). In contrast, humans can solve natural language tasks with only a few illustrative examples(Lake et al., 2015). Motivated by this, demonstration-based learning has been introduced to augment the in- put with a few examples and labels. For instance, Brown et al. (2020) simply picked up to 32 ran- domly sampled instances and directly concatenated them with the input to perform in-context learning with the model frozen and significantly boosted the performance. Lee et al. (2022) concatenated the input with task demonstrations to create augmented input and fed them into PLMs to obtain improved token representations to do sequence labeling in a 28.71 46.25 45.99 25 30 35 40 45 50 No Demonstration Standard Demonstration Random Demonstration X X [SEP] Obama returns to White House. Obama is PER. X [SEP] war House to on Arafat returns F1 Score Figure 1: Performance for different demonstrations on CoNLL03 5-shot support set. Here X denotes the input sentence, such as "Jobs was born in America.", and we show the PER part of the whole demonstration in applegreen for visualization. Surprisingly, random tokens can be good demonstrations too. classifier-based fine-tuning way. However, how and why such demonstrations help remains unclear. As such, there has been a growing amount of work investigating the robust- ness and interpretability of demonstration-based learning. For instance, Lu et al. (2021) reported that few-shot text classification is very sensitive to the ordering of the demonstrations in in-context learning. On a wide range of low-resource Nat- ural Language Understanding (NLU) tasks, Min et al. (2022b) investigated why demonstrations in in-context learning can bring performance …

Content truncated to 500 words