Title:
|
Supporting unsupervised context identification using social and physical sensors
|
Context-awareness is a popular subject in ubiquitous computing, wearable computing and
information retrieval. Location, conditions, activity, task and affect are all viewed as useful elements of context-awareness. Devices equipped with appropriate sensors may collect
contextual information directly through infrastructural services or indirectly, through
inference from observed conditions.
This thesis discusses approaches to representing and sharing such data, such as the
selection, development or use of semantic ontologies able to support context-awareness.
The problem of selecting an appropriate knowledge representation for activities requiring
information about context and the cost of developing, accessing and maintaining ontologies
suggests that an approach that automates some part of this activity would be beneficial.
We suggest that inspiration could and should be drawn from the mechanisms that govern
natural language and classification.
We explore and extend an existing approach to unsupervised labelling of contexts. We
identify failure modes and potential mitigating strategies for an unsupervised approach
depending solely on local interactions. A third source of information in context-awareness
is the social sensor. Information shared on the social web-Flickr, Facebook or Twitter, for
example-is a rich source of localised information both directly and as a data source for
unsupervised learning. We examine data taken from social sensors such as social tagging
services, Facebook and Twitter, identifying significant variation in content and encoding
which we suggest may be of benefit to the primary user populations of the site. We explore
tools intended to support access to and reuse of compound terms. We adapt a well-known
experiment in socially shared cognition to a crowdsourcing platform, finding support
for the view that the context of message production affects message comprehension.
We demonstrate examples of the use of crowdsourced data to ground ontology labels in
observation. Finally, we use data taken from a number of social networks to demonstrate
the tracking of a single concept across time and space, identifying localised variation in
its interpretation and prevalence.
|