SalKG: Learning From Knowledge Graph Explanations for Commonsense Reasoning

Authors: Aaron Chan, Jiashu Xu, Boyuan Long, Soumya Sanyal, Tanishq Gupta, Xiang Ren

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On three commonsense QA benchmarks (CSQA, OBQA, CODAH) and a range of KG-augmented models, we show that SALKG can yield considerable performance gains up to 2.76% absolute improvement on CSQA.
Researcher Affiliation Academia University of Southern California, IIT Delhi {chanaaro, boyuanlo, jiashuxu, soumyasa, xiangren}@usc.edu, Tanishq.Gupta.mt617@maths.iitd.ac.in
Pseudocode No The paper does not contain any sections or figures explicitly labeled 'Pseudocode' or 'Algorithm', nor are there any structured code-like blocks outlining procedures.
Open Source Code Yes Code and data are available at: https://github.com/INK-USC/Sal KG.
Open Datasets Yes We use the CSQA [52] and OBQA [39] multi-choice QA datasets. ... As in prior works, we use the Concept Net [49] KG for both datasets.
Dataset Splits No For CSQA, we use the accepted in-house data split from [31], as the official test labels are not public.
Hardware Specification No The paper does not specify the exact hardware components (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper mentions software components like BERT, RoBERTa, and PyTorch, but it does not specify their version numbers or other ancillary software dependencies with specific versioning for reproducibility.
Experiment Setup Yes We use thresholds T = 0.01 and k = 10 for coarse and fine explanations, respectively. For text encoders, we use BERT(-Base) [11] and Ro BERTa(-Large) [35]. For graph encoders, we use MHGRN [13], Path Gen [56], and Relation Network (RN) [46, 31]. ... LS = Ltask + λLsal, where λ 0 is a loss weighting parameter.