Graph Component Contrastive Learning for Concept Relatedness Estimation

Authors: Yueen Ma, Zixing Song, Xuming Hu, Jingjing Li, Yifei Zhang, Irwin King

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on three datasets show significant improvement over the state-of-the-art model. Detailed ablation studies demonstrate that our proposed approach can effectively capture the high-order relationship among concepts. We conduct comprehensive experiments with three different Transformer models on three datasets.
Researcher Affiliation Academia 1The Chinese University of Hong Kong, 2Tsinghua University
Pseudocode No The paper describes the proposed methods using textual explanations and mathematical equations, but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available on Github1. 1Github: https://github.com/Panmani/GCCL
Open Datasets Yes We use the official dataset split for WORD whose train-test ratio is approximately 2:1. Since CNSE and CNSS do not provide an official dataset split, they are split randomly with a train-dev-test ratio of 7:2:1.
Dataset Splits Yes Since CNSE and CNSS do not provide an official dataset split, they are split randomly with a train-dev-test ratio of 7:2:1.
Hardware Specification Yes Experiments are conducted on four Nvidia TITAN V GPUs.
Software Dependencies No The paper mentions software components like 'Adam W optimizer' and specific 'Transformer models (BERT, RoBERTa, XLNet)', but does not provide version numbers for any libraries or programming languages.
Experiment Setup Yes We use the Adam W optimizer (Loshchilov and Hutter 2019) with learning rate = 1e-5 and ϵ = 1e-8, following a linear schedule. The Transformer models are trained for 5 epochs. For GC-NCE, we use α = 10. For Mo Co (He et al. 2020; Chen et al. 2020b), we use queue size Q = 32, momentum coefficient m = 1 - 1e-4, and temperature τ = 0.1. We use β = 0.1 for the overall loss.