Processing internet-derived text - creating a corpus of usenet messages.

In recent years, linguists have become increasingly interested in the language of the Internet—both as an object of investigation as well as a source of authentic data to complement traditional electronic corpora. However, Internet-derived data is typically very messy data and a conversion process is often required in order to enable researchers to carry out a reliable quantitative investigation of the patterns observed with the help of standard corpus tools. In this article, I discuss the technical and methodological aspects involved in creating a large corpus of asynchronous computer-mediated communication by downloading and post-processing hundreds of thousands messages posted in twelve Usenet newsgroups. After describing how messages can be arranged into hierarchically structured discussion threads, I focus at some length on the strategies that are required to correctly assign authorship to the different textual elements in individual messages. My algorithms have a success rate of well over 90% for most newsgroups and the resulting corpus can thus serve as a suitable basis for an investigation into the interactive strategies employed in this particular type of written communication