GNN is a Counter? Revisiting GNN for Question Answering

Authors: Kuan Wang, Yuyu Zhang, Diyi Yang, Le Song, Tao Qin

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

Reproducibility Variable Result LLM Response
Research Type Experimental We discover that even a very simple graph neural counter can outperform all the existing GNN modules on Commonsense QA and Open Book QA, two popular QA benchmark datasets which heavily rely on knowledge-aware reasoning.
Researcher Affiliation Collaboration 1Georgia Institute of Technology 2Microsoft Research Asia 3Bio Map 4MBZUAI {kuanwang, yuyu, dyang888}@gatech.edu lsong@cc.gatech.edu taoqin@microsoft.com
Pseudocode Yes Algorithm 1 Py Torch-style code of GSC
Open Source Code No The paper does not provide a concrete link or explicit statement about the availability of source code for the described methodology.
Open Datasets Yes We conduct extensive experiments on Commonsense QA (Talmor et al., 2019) and Open Book QA (Mihaylov et al., 2018), two popular QA benchmark datasets that heavily rely on knowledge-aware reasoning capability.
Dataset Splits Yes Hence, following the data split of Lin et al. (2019), we experiment and report the accuracy on the in-house dev (IHdev) and test (IHtest) splits. We also report the accuracy of our final system on the official test set.
Hardware Specification Yes On a single Quadro RTX6000 GPU, each GSC training only takes about 2 hours to converge, while other methods often take 10+ hours.
Software Dependencies No The paper mentions 'Py Torch-style code' and 'RAdam' optimizer but does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We use RAdam (Liu et al., 2020) as the optimizer and set the batch size to 128. The learning rate is 1e-5 for Ro BERTa and 1e-2 for GSC. The maximum number of epoch is set to 30 for Commonsense QA and 75 for Open Book QA.