Pseudo-Label Calibration Semi-supervised Multi-Modal Entity Alignment

Authors: Luyao Wang, Pengnian Qi, Xigang Bao, Chunlai Zhou, Biao Qin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two MMEA datasets demonstrate the effectiveness of our PCMEA, which yields state-of-the-art performance.
Researcher Affiliation Academia School of Information, Renmin University of China {wangluyao123, pengnianqi, baoxigang, czhou, qinbiao} @ruc.edu.cn
Pseudocode No The paper describes its methods and components with equations but does not include explicit pseudocode blocks or algorithm listings.
Open Source Code No The paper mentions that pre-trained language models are downloaded from Hugging Face but does not provide any statement or link for the open-source code of their proposed PCMEA method.
Open Datasets Yes Two cross-KG EA datasets are adopted for evaluation, including FB15K-DB15K and FB15K-YAGO15K, which are the most representative datasets in MMEA task (Chen et al. 2020, 2022a; Lin et al. 2022).
Dataset Splits No The paper mentions "different proportions of entity alignment seeds, i.e., 2:8, 5:5, and 8:2" for training/testing, but does not explicitly specify a separate validation dataset split.
Hardware Specification Yes All experiments were conducted on a server with two GPUs (NVIDIA-SMI 3090).
Software Dependencies No The paper mentions PyTorch and specific pre-trained language models (Bert, T5, RoBERTa, Albert, ChatGLM-6B, LLaMA-7B) but does not provide specific version numbers for PyTorch or the PLMs, only stating "all of them are base version".
Experiment Setup Yes The main hyper-parameters in our method are momentum coefficient κ, momentum network update span ρ, and time of changing training strategy ts. For momentum coefficient κ, a proper large κ (e.g. 0.999) bring better stability and accuracy in Hits@1 and MRR, illustrating momentum-based contrast learning is more effective than just contrast learning. Varying time span ρ shows little difference. For time ts of changing training strategy, time ts obvious effects the surge in Hits@1 after changing the strategy, with ts = 500 allowing faster convergence.