Learning and Leveraging Structured Knowledge from User-Generated Social Media Data



Dong, Hang ORCID: 0000-0001-6828-6891
(2020) Learning and Leveraging Structured Knowledge from User-Generated Social Media Data. PhD thesis, University of Liverpool.

[img] Text
Thesis-HD-final-20-Apr.pdf - Unspecified

Download (3MB) | Preview

Abstract

Knowledge has long been a crucial element in Artificial Intelligence (AI), which can be traced back to knowledge-based systems, or expert systems, in the 1960s. Knowledge provides contexts to facilitate machine understanding and improves the explainability and performance of many semantic-based applications. The acquisition of knowledge is, however, a complex step, normally requiring much effort and time from domain experts. In machine learning as one key domain of AI, the learning and leveraging of structured knowledge, such as ontologies and knowledge graphs, have become popular in recent years with the advent of massive user-generated social media data. The main hypothesis in this thesis is therefore that a substantial amount of useful knowledge can be derived from user-generated social media data. A popular, common type of social media data is social tagging data, accumulated from users' tagging in social media platforms. Social tagging data exhibit unstructured characteristics, including noisiness, flatness, sparsity, incompleteness, which prevent their efficient knowledge discovery and usage. The aim of this thesis is thus to learn useful structured knowledge from social media data regarding these unstructured characteristics. Several research questions have then been formulated related to the hypothesis and the research challenges. A knowledge-centred view has been considered throughout this thesis: knowledge bridges the gap between massive user-generated data to semantic-based applications. The study first reviews concepts related to structured knowledge, then focuses on two main parts, learning structured knowledge and leveraging structured knowledge from social tagging data. To learn structured knowledge, a machine learning system is proposed to predict subsumption relations from social tags. The main idea is to learn to predict accurate relations with features, generated with probabilistic topic modelling and founded on a formal set of assumptions on deriving subsumption relations. Tag concept hierarchies can then be organised to enrich existing Knowledge Bases (KBs), such as DBpedia and ACM Computing Classification Systems. The study presents relation-level evaluation, ontology-level evaluation, and the novel, Knowledge Base Enrichment based evaluation, and shows that the proposed approach can generate high quality and meaningful hierarchies to enrich existing KBs. To leverage structured knowledge of tags, the research focuses on the task of automated social annotation and propose a knowledge-enhanced deep learning model. Semantic-based loss regularisation has been proposed to enhance the deep learning model with the similarity and subsumption relations between tags. Besides, a novel, guided attention mechanism, has been proposed to mimic the users' behaviour of reading the title before digesting the content for annotation. The integrated model, Joint Multi-label Attention Network (JMAN), significantly outperformed the state-of-the-art, popular baseline methods, with consistent performance gain of the semantic-based loss regularisers on several deep learning models, on four real-world datasets. With the careful treatment of the unstructured characteristics and with the novel probabilistic and neural network based approaches, useful knowledge can be learned from user-generated social media data and leveraged to support semantic-based applications. This validates the hypothesis of the research and addresses the research questions. Future studies are considered to explore methods to efficiently learn and leverage other various types of structured knowledge and to extend current approaches to other user-generated data.

Item Type: Thesis (PhD)
Uncontrolled Keywords: Knowledge Engineering, Social Media Data, Folksonomies, Social Tags, Ontology, Probabilistic Topic Models, Deep Learning, Neural Networks, Attention Mechanisms, User-Generated Data, Relation Extraction, Knowledge Base Enrichment, Multi-Label Classification, Social Annotation
Divisions: Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science
Depositing User: Symplectic Admin
Date Deposited: 18 Aug 2020 09:56
Last Modified: 18 Jan 2023 23:54
DOI: 10.17638/03084182
Supervisors:
URI: https://livrepository.liverpool.ac.uk/id/eprint/3084182