Automatic refinement of large-scale cross-domain knowledge graphs


Melo, André


[img]
Preview
PDF
thesis.pdf - Published

Download (25MB)

URL: https://ub-madoc.bib.uni-mannheim.de/45980
URN: urn:nbn:de:bsz:180-madoc-459801
Document Type: Doctoral dissertation
Year of publication: 2018
Place of publication: Mannheim
University: Universität Mannheim
Evaluator: Paulheim, Heiko
Date of oral examination: 24 August 2018
Publication language: English
Institution: School of Business Informatics and Mathematics > Data Science (Paulheim 2018-)
License: CC BY 4.0 Creative Commons Attribution 4.0 International (CC BY 4.0)
Subject: 004 Computer science, internet
Subject headings (SWD): Semantic Web
Keywords (English): knowledge graph refinement , machine learning , semantic web
Abstract: Knowledge graphs are a way to represent complex structured and unstructured information integrated into an ontology, with which one can reason about the existing information to deduce new information or highlight inconsistencies. Knowledge graphs are divided into the terminology box (TBox), also known as ontology, and the assertions box (ABox). The former consists of a set of schema axioms defining classes and properties which describe the data domain. Whereas the ABox consists of a set of facts describing instances in terms of the TBox vocabulary. In the recent years, there have been several initiatives for creating large-scale cross-domain knowledge graphs, both free and commercial, with DBpedia, YAGO, and Wikidata being amongst the most successful free datasets. Those graphs are often constructed with the extraction of information from semi-structured knowledge, such as Wikipedia, or unstructured text from the web using NLP methods. It is unlikely, in particular when heuristic methods are applied and unreliable sources are used, that the knowledge graph is fully correct or complete. There is a tradeoff between completeness and correctness, which is addressed differently in each knowledge graph’s construction approach. There is a wide variety of applications for knowledge graphs, e.g. semantic search and discovery, question answering, recommender systems, expert systems and personal assistants. The quality of a knowledge graph is crucial for its applications. In order to further increase the quality of such large-scale knowledge graphs, various automatic refinement methods have been proposed. Those methods try to infer and add missing knowledge to the graph, or detect erroneous pieces of information. In this thesis, we investigate the problem of automatic knowledge graph refinement and propose methods that address the problem from two directions, automatic refinement of the TBox and of the ABox. In Part I we address the ABox refinement problem. We propose a method for predicting missing type assertions using hierarchical multilabel classifiers and ingoing/ outgoing links as features. We also present an approach to detection of relation assertion errors which exploits type and path patterns in the graph. Moreover, we propose an approach to correction of relation errors originating from confusions between entities. Also in the ABox refinement direction, we propose a knowledge graph model and process for synthesizing knowledge graphs for benchmarking ABox completion methods. In Part II we address the TBox refinement problem. We propose methods for inducing flexible relation constraints from the ABox, which are expressed using SHACL.We introduce an ILP refinement step which exploits correlations between numerical attributes and relations in order to the efficiently learn Horn rules with numerical attributes. Finally, we investigate the introduction of lexical information from textual corpora into the ILP algorithm in order to improve quality of induced class expressions.
Additional information: Vollständiger Name: André de Oliveira Melo




Dieser Eintrag ist Teil der Universitätsbibliographie.

Das Dokument wird vom Publikationsserver der Universitätsbibliothek Mannheim bereitgestellt.




Metadata export


Citation


+ Search Authors in

BASE: Melo, André

Google Scholar: Melo, André

+ Download Statistics

Downloads per month over past year

View more statistics



You have found an error? Please let us know about your desired correction here: E-Mail


Actions (login required)

Show item Show item