Panda Alert Research Proposal

点击翻译按钮获取中文摘要

The Panda Alert System aims to improve the state of the art in early disease outbreakdetection by incorporating linguistic features into the language models which werepreviously not known/seen to indicate any risk to human health. The project willdevelop a fully functioning, real-time alerting and mapping system.This kind of research project assumes significant knowledge/use of softwareengineering methodologies combined with novel research into unsupervised orsemi-supervised NLP models. We propose a layered, modular architecturedetachable from the core NLP engine, which should demonstrate a universaldetection capability easily transferable to new domains.The most likely approach would be a semi-supervised, bootstrapping model, whichlearns from a small amount of training data to generalise over the unknown domain.Harnessing the latest NLP machine learning methods such as neural networks (deeplearning), the model analyses news reports to flag any risks to public health. All toolsand data sets from this research will be open sourced and available for downloadfrom public repositories such as GitHub or Google Code.It is vital for the future reusability of the core NLP engine to be domain-agnostic sothat it can be extended and adapted via a clear programming interface to new eventdetection tasks. The real-time analysis system will have to handle hundreds ofthousands of news articles and social media items per day in multiple languages.The research part of the project will take the state of the art in NLP and AI andextend it to establish novel methods of NLP general discourse pattern detection inorder to analyse statements for known and unknown features potentially indicativeof the fact-bearing target knowledge. The new technique should be capable ofhandling the identification and reporting of other target knowledge/facts from casualonline user/news activity with minimum or no training.In year one, I shall be mainly focusing on toponym resolution and geoparsing. Thesetechniques comprise NLP techniques for identification of place names in text andtheir subsequent resolution to geographical coordinates. This task is easy for humansto perform, however for a machine to tell which London (UK, Canada, etc.) wasmentioned in a text is a non-trivial job. I aim to improve on the existing baselines byresearching a novel method of geoparsing. We aim to publish the findings next yearat a respected conference.