Tkachenko, Nataliya (2019) Using semantic drift on social media for event detection, differentiation and segmentation. PhD thesis, University of Warwick.

Preview

PDF WRAP_Theses_Tkachenko_2019.pdf - Submitted Version - Requires a PDF viewer. Download (28Mb) | Preview

Request Changes to record.

Abstract

With observable paradigm shift in computer science from predictive modeling to the generative one, it became important to maximise exploration of the pathways towards useful data production. With currently dominating statistical and compositional data augmentation strategies, opportunities also emerged for more application-driven routes. The main value of such approaches lies in their capacity to offer insights into context or event specific data productions, currently overlooked by more topologically neutral machine learning approaches. The purpose of this thesis is therefore to provide empirical evidence for useful data generation by dynamic event-specific lexical semantic resources.

Various Web 2.0 applications due to their popularity have been accumulating large amounts of semantically rich metadata, which became readily available and easily exploitable. Tags, usually consisting of a single word, are one type of such data. Tag uses can vary largely across systems and platforms; Also known under the term folksonomy, tags are usually non-hierarchical and open-ended, thus re-flecting users' unique perspectives regarding various contexts, or resources. This platform-enabled liberty of expression, however, has led to situations of frequent semantic ambiguity due to spelling mistakes, morphological variations, polysemy, multilingualism or inaccurate tag-to-resource associations. As a consequence, tag spaces are often regarded as inconsistent, noisy and hardly reliable data sources.

Recent surge of interest amongst distributional semanticists in long- and short-term fluctuations of word meanings on social media has suggested routes for successful temporal sense disambiguation, thus inviting discussions around useful real-world applications for such emerging data resources. One of such applications - event analytics from the crowd behaviour perspective - is gaining an increasing attention from researchers and practitioners, especially in the fields of operations and situational management. Pursuing pragmatic aims of event detection, differen- tiation and segmentation, this application domain is represented predominantly by repetitive catastrophic events (such as natural hazards), during which directly or indirectly exposed populations tend to share their situational experiences on social media.

This thesis consists of three main parts, each corresponding to specific problem in event analytics: (i) detection, (ii) differentiation and (iii) segmentation. In the first part I used the concept of ontological semantic proximity on the words candidates for semantic drift in order to highlight the dynamics of their semantic oscillations within event-specific category (i.e., flooding). In my second experiment I followed on these initial findings and performed an analysis verifying whether semantically unstable lexical material can augment our knowledge about main sub-types of floods, such as `slow' (e.g., groundwater and pluvial floods) and `fast' (surface water and riverine floods) ones. In my third experiment I employed combined lexico-visual modalities of the crowdsourced material to reconstruct changing perceptions of flood events in order to understand how event severity can or cannot determine situationally resilient behaviours.

Item Type:	Thesis or Dissertation (PhD)
Subjects:	G Geography. Anthropology. Recreation > GB Physical geography Q Science > QA Mathematics
Library of Congress Subject Headings (LCSH):	Data mining -- Online social networks, User-generated content, Social media, Metadata, Natural disaster warning systems, Floods
Official Date:	June 2019
Dates:	Date Event June 2019 UNSPECIFIED
Institution:	University of Warwick
Theses Department:	Department of Computer Science
Thesis Type:	PhD
Publication Status:	Unpublished
Supervisor(s)/Advisor:	Jarvis, Stephen A., 1970- ; Procter, Rob
Format of File:	pdf
Extent:	xviii, 150 leaves : illustrations, charts
Language:	eng

Request changes or add full text files to a record

Repository staff actions (login required)

View Item

Downloads