Decoupling Knowledge from Memorization: Retrieval-augmented Prompt Learning

Authors: Xiang Chen, Lei Li, Ningyu Zhang, Xiaozhuan Liang, Shumin Deng, Chuanqi Tan, Fei Huang, Luo Si, Huajun Chen

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that RETROPROMPT can obtain better performance in both few-shot and zero-shot settings.
Researcher Affiliation Collaboration Xiang Chen1,2 , Lei Li1,2 , Ningyu Zhang1,2 , Xiaozhuan Liang1,2, Shumin Deng1,2, Chuanqi Tan3, Fei Huang3, Luo Si3, Huajun Chen1,2 1Zhejiang University & AZFT Joint Lab for Knowledge Engine, China 2Hangzhou Innovation Center, Zhejiang University, China 3Alibaba Group, China
Pseudocode No No explicit pseudocode or algorithm blocks found.
Open Source Code Yes Code is available in https://github.com/zjunlp/Prompt KG/tree/main/research/ Retro Prompt.
Open Datasets Yes We evaluate RETROPROMPT on several types of natural language understanding tasks, including single sentence classification tasks (SST-2 [54], MR [43], and CR [19]) and sentence pair classification tasks (MNLI [58], QNLI [46], and QQP5). To further evaluate the effectiveness of the proposed approach with multi-class classification, we also conduct experiments on the information extraction tasks, including Few NERD [8], Sem Eval 2010 Task 8 (Sem Eval) [17], and TACRED [61].
Dataset Splits No We follow the few-shot setting of LM-BFF [12] to conduct 4-shot and 16-shot experiments and evaluate the average performance with a fixed set of seeds, Sseed, across several different sampled Dtrain for each task. No explicit train/validation/test dataset splits (percentages or counts) are provided in the main text for the overall datasets beyond the few-shot setting description.
Hardware Specification Yes The experiments are implemented on 1 NVIDIA V100 and utilize Pytorch [44] as the base library.
Software Dependencies No The paper mentions software like Pytorch [44], Ro BERTalarge [38], and FAISS [21], but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We list the specific experimental settings and tuning retrieve parameters in Appendix C and D.