The Best Kept Secrets with Corpus Linguistics

This paper presents the use of corpus linguistics techniques on supposedly “clean” corpora and identifies potential pitfalls. Our work relates to the task of filtering sensitive content, in which data security is strategically important for the protection of government and military information, and of growing importance in combating identity fraud. A naïve keyword filtering approach produces a large proportion of false positives, and the need for more fine-grained approaches, suggests the consideration of using corpus linguistics for such content filtering. We present work undertaken on the Enron corpus, a collection of emails that has had various tasks undertaken on various portions and versions of it by other researchers. We have made some efforts to reconcile differences between the different versions by considering what impact some of these versions have on our results. Our anticipated efforts in using automatic ontology learning [Gillam, Tariq and Ahmad 2005; Gillam and Ahmad 2005], and local grammars [Ahmad, Gillam and Cheng 2006] to discover points of interest within the Enron corpus have been reoriented to the problem of discovering “confidentiality banners” with a view to removing them from the collection. Our work to date makes strong use of collocation patterns [Smada 1993] to identify the signatures of these banners, and to use the banners themselves to better understand the different versions of the Enron corpus. We further consider the use of extended collocation patterns to identify text “zones”, following [Teufel and Moens 2000], and the subsequent potential for sentiment analysis [Klimt and Yang 2004]; [Pang, Lee and Vaithyanathan 2002]