Online Noisy Continual Relation Learning

Authors: Guozheng Li, Peng Wang, Qiqing Luo, Yanhe Liu, Wenjun Ke

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on Few Rel, TACRED and NYT-H with real-world noise demonstrate that our framework greatly outperforms the combinations of the state-of-the-art online continual learning and noisy label learning methods.
Researcher Affiliation Academia Guozheng Li1, Peng Wang1*, Qiqing Luo1, Yanhe Liu1, Wenjun Ke1,2 1School of Computer Science and Engineering, Southeast University 2Beijing Institute of Computer Technology and Application {gzli, pwang, qqluo, liuyanhe}@seu.edu.cn, kewenjun2191@163.com
Pseudocode Yes Algorithm 1: Online Purifying and Learning for S6
Open Source Code No The paper does not contain any explicit statement about releasing source code or provide a link to a code repository.
Open Datasets Yes We test on two datasets of Few Rel (Han et al. 2018) and TACRED (Zhang et al. 2017) with synthetic noise. And we adopt the NYT-H (Zhu et al. 2020) dataset obtained by distant supervision (Mintz et al. 2009) to evaluate under the real-world noise.
Dataset Splits No The paper describes how tasks are created and mentions a test set for NYT-H, but does not explicitly specify train/validation/test dataset splits with percentages or counts for reproducibility, focusing instead on online streaming data and replay buffers.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or cloud instance types used for experiments.
Software Dependencies No The paper mentions using BERT as a backbone but does not provide specific version numbers for software dependencies like BERT, PyTorch/TensorFlow, Python, or other libraries.
Experiment Setup Yes The size of memory for replaying is set to 1,600, 800, 400 for Few Rel, TACRED and NYT-H, respectively. For noisy label learning methods, we use α = β = 1.0 in SL, α = 0.1, β = 0.4, λ = 600 in Pencil, and λ = 0.1 in Jo Co R, respectively. For fair comparison, We need to ensure that S6 uses roughly the same size of replay buffer as others. For Few Rel, we set the size of D, C, N as 1,000, 1,000 and 2,000. For TACRED and NYT-H, we set the size of D, C, N as 200, 200 and 400. The learning rates of fϕ( ) and fθ( ) are 5e-6 and 5e-5. The update sizes K1 and K2 are both 5. The batch size is 16 for both online and offline training. And we train 60 epochs in the finetuning stage.