An ontological analysis of vague motion verbs, with an application to event recognition

This research presents a methodology for the ontological formalisation of
vague spatial concepts from natural language, with an application to the
automatic recognition of event occurrences on video data. The main issue
faced when defining concepts sourced from language is vagueness,
related to the presence of ambiguities and borderline cases even in simple
concepts such as ‘near’, ‘fast’, ‘big’, etc. Other issues specific to this
semantic domain are saliency, granularity and uncertainty.
In this work, the issue of vagueness in formal semantics is discussed
and a methodology based on supervaluation semantics is proposed. This
constitutes the basis for the formalisation of an ontology of vague spatial
concepts based on classical logic, Event Calculus and supervaluation
semantics. This ontology is structured in layers where high-level concepts,
corresponding to complex actions and events, are inferred through
mid-level concepts, corresponding to simple processes and properties of
objects, and low-level primitive concepts, representing the most essential
spatio-temporal characteristics of the real world.
The development of ProVision, an event recognition system based on a
logic-programming implementation of the ontology, demonstrates a practical
application of the methodology. ProVision grounds the ontology on
data representing the content of simple video scenes, leading to the inference
of event occurrences and other high-level concepts.
The contribution of this research is a methodology for the semantic
characterisation of vague and qualitative concepts. This methodology addresses
the issue of vagueness in ontologies and demonstrates the applicability
of a supervaluationist approach to the formalisation of vague concepts.
It is also proven to be effective towards solving a practical reasoning
task, such as the event recognition on which this work focuses.