RoPDA: Robust Prompt-Based Data Augmentation for Low-Resource Named Entity Recognition

Authors: Sihan Song, Furao Shen, Jian Zhao

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three popular benchmarks from different domains demonstrate that Ro PDA significantly improves upon strong baselines, and also outperforms state-of-the-art semi-supervised learning methods when unlabeled data is included.
Researcher Affiliation Academia Sihan Song1,2, Furao Shen1,3*, Jian Zhao1,4 1 National Key Laboratory for Novel Software Technology, Nanjing University, China 2Department of Computer Science and Technology, Nanjing University, China 3School of Artificial Intelligence, Nanjing University, China 4School of Electronic Science and Engineering, Nanjing University, China whalesihan@smail.nju.edu.cn, frshen@nju.edu.cn, jianzhao@nju.edu.cn
Pseudocode No The paper describes the methodology using textual explanations and diagrams (Figure 2), but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about releasing source code for the described methodology, nor does it include a link to a code repository.
Open Datasets Yes Datasets We conduct experiments on three datasets. (a) Co NLL03 (Tjong Kim Sang and De Meulder 2003) is a collection of news wire articles from the Reuters Corpus, containing 4 entity types. (b) MIT Restaurant (Liu et al. 2013) is a collection of user utterances in the restaurant domain with 8 entity types. (c) MIT Movie (Liu et al. 2013) consists of user utterances in the movie domain with 12 entity types.
Dataset Splits No We create four low-resource settings of shot-5/10/20/50 for each dataset. In the shot-K setting, we sample K samples from the Train set for each entity type as the training set and add the remaining to the unlabeled set.
Hardware Specification No The paper does not provide any specific details regarding the hardware used for running the experiments (e.g., specific GPU models, CPU models, or cloud infrastructure).
Software Dependencies No The paper mentions using the T5 model as the backbone PLM ('we choose the T5 (Raffel et al. 2020) model as our backbone sequence-to-sequence PLM'), but does not provide specific version numbers for T5 or any other software dependencies, libraries, or frameworks used.
Experiment Setup Yes we compare the soft prompt model with the standard T5 model, which undergoes fine-tuning using a learning rate of 1e-4. ... K is initially set to 1, ... with M=2 and N=3 providing the best results. In addition, the best performance is attained when M and N are both randomly chosen from {1, 2, 3}.