Abstract
dc:descriptionThe exponential growth of RDF Knowledge Bases has resulted in massive volumes of data withvarying degrees of quality and complexity. Many datasets now contain millions or even billions oftriples, leading to challenges in data representation, extended search times, and difficulties in datacomprehension. These issues underscore the need for tools that can efficiently extract and highlightthe most pertinent information. In this direction, semantic summarization methods have emergedtrying to extract meaning from data while reducing the size of the original graph. Semantic summariescan be exploited instead of the original graph to perform certain tasks more efficiently or effectively,however constructing semantic summaries poses challenges regarding efficiency and quality whereasthere is no widely accepted solution in the domain.In this dissertation, we present novel semantic summarization techniques focusing on both theefficiency of the summarization algorithm as well on the quality of the constructed summary.We initially present SumMER, focusing on the generation of structural, non-quotient semanticsummaries. SumMER moves beyond past approaches on structural summarization, thatselect a single(or just a few) centrality measure(s) for node selection, learning using machine learning techniqueshow to optimally combine multiple centrality measures for node selection. Combining multiplecentrality measures provides a complementary view of the node’s importance, significantly upliftingthe quality of the generated summaries and dominating all other baselines in the domain.Although node selection exploiting multiple centrality measures results in high-qualitysummaries, it faces scalability problems for large graphs. In addition, edges’ importance was notconsidered at all in the final summaries. To overcome those problems we next focus on exploitingembeddings for node and edges representation and selection. This approach combines graphembeddings with machine learning to identify the most significant nodes and edges while employscarefully chosen approximate algorithms for node linking, considering not just the minimum numberof nodes like previous works did, but the importance of edges as well. The resulting algorithms forsummary generation demonstrate high efficiency and reduced execution time being orders ofmagnitude faster than competitors, whereas the generated summaries also enjoy better quality.Finally, instead of focusing on summarizing the entire KB, we next focus on specific parts of thegraph that users might want to summarize focusing on Entity Summarization. We propose EntitySum,an entity summarization approach that ensures data filtering and redundancy reduction, in constrastwith the state-of-the art approaches, which the most of them do not filtering data and struggle tominimize redundancy. In addition our approach, concurrently selecting the most significant triplesbased on centrality measures for object importance and frequency for properties. Our approachfurther incorporates large language models (LLMs) to transform summaries into natural language,enhancing interpretability. We show that EntitySum demonstrates superior efficiency compared tothe state-of-the-art methodologies, whereas it produces consistently better results.
Degree
thesis:*- Grantor dc:publisher
- Ελληνικό Μεσογειακό Πανεπιστήμιο
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Trouli, Georgia-Eirini
- Τρούλη, Γεωργία-Ειρήνη
Subjects
dc:subject × 13- RDF KGs
- Σημασιολογικές συνόψεις
- Συνόψεις γράφων
- Συνόψεις οντοτήτων
- Semantic summaries
- Graph summaries
- Entity summaries
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Τεχνολογία μέσων
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Media Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58019