Relatedness and TBox-Driven Rule Learning in Large Knowledge Bases

Authors: Giuseppe Pirrò2975-2982

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We considered four-real world datasets, that is, WN-18RR and FB15-237 used in (Meilicke et al. 2019), and excerpts of Yago3-10 (Yago) (Gal arraga et al. 2015) and DBpedia (Shiralkar et al. 2017). For all these datasets we used the portion of the TBox schema including subclass information and domain and range of properties. Details about the datasets are available in Table 1. Following (Meilicke et al. 2019), we computed the filtered hits@1, filtered hits@10, and the mean reciprocal rank (MRR); we did not compute the filtered MRR as we are only interested in computing top-k ranks only.
Researcher Affiliation Academia 1Department of Computer Science, Sapienza University of Rome Via Salaria 113, 00198, Rome, Italy pirro@di.uniroma1.it
Pseudocode Yes Algorithm 1: RARL (p, k, d, G, GS, MR, V +, V , rel R, α, β) ... Algorithm 2: generate Candidate Bodies(p, k, d, GS, MR) ... Algorithm 3: get Reduced ABox Graph(G, Πi, V +, V )
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a code repository for RARL.
Open Datasets Yes We considered four-real world datasets, that is, WN-18RR and FB15-237 used in (Meilicke et al. 2019), and excerpts of Yago3-10 (Yago) (Gal arraga et al. 2015) and DBpedia (Shiralkar et al. 2017).
Dataset Splits No Table 1: Datasets characteristics. ... Testset 3K 20K 5K
Hardware Specification Yes We implemented RARL in Java and ran experiments on a laptop with 4 cores (each with 2,7 GHz) and 16GB RAM.
Software Dependencies No We implemented RARL in Java and ran experiments on a laptop with 4 cores (each with 2,7 GHz) and 16GB RAM.
Experiment Setup Yes We considered the following default parameter values: d=3 (max. body length), top Ps=10 (top-10 related predicates), top C=80% (percentage of candidate bodies for which we want to compute confidence), α=β=0.5 (weights for the confidence score in equation (2)), n Exs=80% (number examples used as a percentage of all available positive facts for a predicate).