LEGO: Latent Execution-Guided Reasoning for Multi-Hop Question Answering on Knowledge Graphs

Authors: Hongyu Ren, Hanjun Dai, Bo Dai, Xinyun Chen, Michihiro Yasunaga, Haitian Sun, Dale Schuurmans, Jure Leskovec, Denny Zhou

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on several KGQA benchmarks demonstrate the effectiveness of our framework compared with previous state of the art.
Researcher Affiliation Collaboration 1Stanford University 2Google Brain 3UC Berkeley 4Carnegie Mellon University.
Pseudocode Yes Algorithm 1 Latent Execution-Guided Reasoning (Training) Algorithm 2 Latent Execution-Guided Reasoning (Inference)
Open Source Code Yes The implementation of LEGO can be found in http://github.com/snap-stanford/lego.
Open Datasets Yes We evaluate LEGO on three large-scale multi-hop KGQA benchmark datasets: Meta QA (Zhang et al., 2018), Web Questions SP (WQSP) (Yih et al., 2015) and Complex Web Questions (CWQ) (Talmor & Berant, 2018).
Dataset Splits Yes Table 2. Statistics of the three datasets. Train Dev Test Meta QA-1hop 96,106 9,992 9,947 Meta QA-2hop 118,980 14,872 14,872 Meta QA-3hop 114,196 14,274 14,274 WQSP 2,848 250 1,639 CWQ 27,623 3,518 3,531
Hardware Specification No The paper does not specify the hardware used for experiments.
Software Dependencies No The paper mentions using a pretrained language model (Devlin et al., 2019; Reimers & Gurevych, 2019) and Query2box (Q2B) (Ren et al., 2020) but does not provide specific version numbers for any software dependencies.
Experiment Setup No The paper does not include specific details on hyperparameters or training settings in the main text.