17 Oct 2018 | 5pm | Dr Stuart Middleton

Note: The PDF of Dr Middleton's lecture can be accessed here:

ucl-2018-middleton-v3.pdf

ucl-2018-middleton-v3.pdf

Real-time geoparsing and information extraction from social media streams, online marketplaces and online forums is providing a new 'virtual sensor' capability for open source intelligence (OSINT). End users such as emergency response agencies (e.g. Tsunami early warning centres, Civil protection authorities), news agencies (e.g. Deutsche Welle, BBC News) and law enforcement agencies (e.g. UK Border Force, National Crime Agency) are all actively investigating how best to use this new data to support their day to day requirements.

Research challenges in this area revolve around scalability (i.e. volume of data), trust (i.e. veracity of the content) and responding to the dynamic nature of the content (i.e. velocity & variety of emerging topics). Social media data volumes are often large, with GBytes of text easily obtainable in near-real-time. Algorithm scalability is critical for near-real-time processing, so efficient indexing and/or naive parallelization is important. In the online world fake news is rife so analysing the provenance of sources is very important for verification, along with cross-checking contextual information around factual claims. Natural language processing (NLP) and open information extraction (OpenIE) approaches typically work with large corpus of example posts, or hand crafted language patterns and/or vocabularies. In fast moving events (e.g. breaking news) locations or topics of interest emerge and evolve in real-time, so training data is limited or non-existent.
In this seminar I will present work from three projects (TRIDEC, REVEAL, FloraGuard), outlining the algorithms we designed for near real-time geoparsing and information extraction, and the geospatial information systems (GIS) that used them to provide decision support. I will show the results we obtained, and outlines some lessons learnt along the way.