Extraction of structured semanti knowledge through data mining over social media

• Autores: Daniel Fernández Álvarez
• Directores de la Tesis: Daniel Gayo Avello (dir. tes.) , José Emilio Labra Gayo (dir. tes.)
• Lectura: En la Universidad de Oviedo ( España ) en 2022
• Idioma: español
• Tribunal Calificador de la Tesis: Francisco Ortín Soler (presid.) , Daniel Fernández Lanvín (secret.) , Leyla Jael Garcia Castro (voc.)
• Programa de doctorado: Programa de Doctorado en Informática por la Universidad de Oviedo
• Menciones:
  • Internacional
• Calificación: Sobresaliente Cum Laude
• Enlaces
  • Tesis en acceso abierto en: RUO
• Resumen

  • The first motivation of this thesis is providing a system able to extract RDF shapes by mining text content expressed in social media. Shapes obtained with such a system are a formal and structured expression of conceptual knowledge associated with a certain context. Besides being helpful for text validation and description, those shapes can be used to perform further actions that can be beneficial for the mined source, such as automatic content classification, content suggestion, or automatic generation of forms and templates. While developing such a system, we produced three main contributions: the SEITMA architecture, the sheXer system, and the ClassRank algorithm. SEITMA is an architecture designed for extracting shapes from natural language that combines two main processes: the extraction of RDF from natural language, and the extraction of shapes by mining RDF triples. SEITMA is based on example generalization: it uses examples of texts with instance-level knowledge to produce RDF shapes. We have implemented two prototypes of SEITMA that use different strategies for producing RDF from natural language. We used both prototypes to extract shapes from Wikipedia abstracts. Such abstracts are related to instances of important classes in DBpedia, and each shape generated contains Wikipedia knowledge associated with one of those classes. Our prototypes generate different types of shapes, but they both produce promising results that prove the feasibility of our proposal. sheXer is a system for the automatic extraction of shapes by mining RDF graphs. In our thesis context, sheXer is used by both SEITMA prototypes. However, the specific problems tackled by sheXer are relevant for the Semantic Web community. Even if there are some other approaches for the automatic extraction of shapes from RDF content, sheXer is a competitive alternative, as it has a unique combination of features among existing alternatives: 1) generation of both ShEx and SHACL content, 2) capacity to process large graphs, 3) production of shapes with inverse paths, and 4) shape-interlinkage, i.e., production of shapes which can reference other shapes. sheXer uses an iterative strategy that avoids loading the whole target graph in main memory. We evaluated sheXer's performance w.r.t. execution time and memory usage. We detected that the number of instances relevant for the process is the parameter that affects the performance the most. However, we proved that sheXer produces precise results by using only representative instance samples. ClassRank allows for detecting important classes in RDF graphs. ClassRank assigns each class an importance score w.r.t. the accumulated importance of its instances. The importance at instance-level is calculated using PageRank. We used ClassRank for class priorization in several stages of the experiments with SEITMA prototypes. However, discovering the most important classes in RDF graphs is a problem relevant out of the context of our thesis. We have evaluated ClassRank by comparing the importance rankings that it produces with rankings based on class-usage in SPARQL logs. This evaluation was performed using Wikidata and DBpedia. We compared ClassRank with many other state-of-the-art approaches and demonstrated that our proposal is the one which produces the most similar results to the class-usage rankings. The source code of ClassRank, sheXer, and the SEITMA prototypes has been publicly released, and every experiment described in this document can be replicated.