Marrying Uncertainty and Time in Knowledge Graphs

Authors: Melisachew Chekol, Giuseppe Pirr˜, Joerg Schoenfisch, Heiner Stuckenschmidt

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We report on an experimental evaluation comparing the MLN and PSL approaches. We carry out a set of experiments using state-of-the-art MLN solvers and their scalable variants. We conducted two different kinds of experiments: (i) performance test in terms of running times for MAP inference comparing three state-of-the-art solvers, and (ii) conflict detection in a noisy setting.
Researcher Affiliation Academia Melisachew Wudage Chekol Data and Web Science Group, University of Mannheim mel@informatik.uni-mannheim.de Giuseppe Pirr o Rende(CS), Italy pirro@icar.cnr.it Joerg Schoenfisch Data and Web Science Group University of Mannheim joerg@dwslab.de Heiner Stuckenschmidt Data and Web Science Group University of Mannheim heiner@informatik.uni-mannheim.de
Pseudocode Yes Figure 1: A set of temporal RDF inference rules that we denote by F. check(T1, T2) = false if T1 T2 = and true otherwise. α denotes the RDF type relation and q is a shorthand for quad. Moreover, all of the formulas are universally quantified over all the variables.
Open Source Code No The paper mentions using and implementing extensions for third-party tools (Tuffy, nRockIt, PSL, nPSL) but does not state that their specific implementations or modifications are open-source or provide a link to their code.
Open Datasets Yes We experimented with YAGO (Gal arraga et al. 2015) to mine temporal rules as discussed below. Mined: With a workaround to AIME, we were able to learn rules of the following form from the YAGO dataset.
Dataset Splits No The paper describes how erroneous facts were injected into datasets and how weights were assigned, but it does not specify train, validation, or test dataset splits.
Hardware Specification Yes We ran the experiments on a 2GHz 24-core processor with 386GB of RAM running Debian 8.
Software Dependencies No The paper mentions the tools used (Tuffy, n Rock It, PSL) and cites papers for them, but it does not provide specific version numbers for these software components.
Experiment Setup Yes We generated erroneous temporal facts specifying the plays For, birthdate, and deathdate relations. We injected a fraction of 10%, 25%, 50%, 75%, and 100% incorrect facts to the Wikidata dataset. For instance, injecting 25% erroneous facts means that we added 25% additional wrong facts for each of the three relations. We randomly assigned weights in the range [0.5, 1.0] to the newly added facts and [0.8, 1.0] to each of the original temporal facts. We repeated each experiment 10 times and present average scores.