Our research group has been awarded funding by the National Cancer Institute for a rhetorical analysis of “deception” in the Tobacco Documents (TDs). These documents, which were released by tobacco industry defendants as a result of state and federal litigation and legislative hearings, cover the complete range of corporate operations in the tobacco companies, from memos to research papers to procurement invoices. The documents are stored physically in depositories in Minneapolis (the site of the original trial) and Guildford, England, and large collections of them (more than five million documents) are now available in electronic form on the Web as well. The documents represent a rich source of corporate and technical discourse which had never been subjected to systematic linguistic analysis; indeed, we are not aware that any similar corporate body of documents has ever been available for analysis. Rather than choosing specific documents for analysis, a method which would leave itself open to attack on grounds of highly selective use of data, the premise of our work is to treat the TDs as a corpus, and to apply accepted methods of corpus and forensic linguistics and rhetorical analysis. Of course, this required that we create sub-corpora for study, since we do not have the resources to include the entire set. Here we will present our experience with the planning and creation of our TD corpora: the sampling strategy, archiving, retrieval, and ultimately, making the corpora available via the Internet for further research. Our initial goal was to create a series of corpora from the TDs in order to 1) Identify TDs in which rhetorical manipulation (“deception”) may have occurred, and to estimate the extent and prevalence of manipulation; and 2) Analyze manipulation we find in order to classify it and develop means to identify similar manipulation in other industrial situations. To do so we have followed a three-part strategy for corpus creation which emphasizes rigorous sampling methods. We first drew a limited sample from the entire body of TDs so that we could determine the best classification of text types and estimate their proportions within the overall body of texts. From those text types which we considered relevant to (i.e. subject to) rhetorical manipulation, we devised quotas for creating a reference corpus of approximately 500,000 words, which we estimated to consist of 808 documents. For this reference corpus, all relevant TDs were sampled whether or not they were thought to contain any manipulation. Finally, we are presently compiling a corpus which includes all texts which we determine to contain any rhetorical manipulation, along with parallel corpora of earlier drafts of the same texts or versions of the same texts prepared for other audiences, so that detailed analysis of rhetorical manipulation can be carried out for itself and by comparison with cross-draft and cross-audience TDs. As it has turned out, the plan has been effective in the first two parts which we have now completed, but we have had to make adjustments at several points in order to take account of our preliminary findings. Once we began the process of collecting documents we immediately encountered two problems related to archiving and processing the data. The first is that there was no text available. Rather than being stored digitally as the plain ASCII text which we needed for computer-assisted corpus analysis, the tobacco documents are stored as image files, usually as TIF type. This problem was compounded by the fact that the images, although stored as large high-resolution files, are generally too poor in quality for automated text processing such as scanning and OCR. They often have pages that are tipped to one side or, in the case of dot-matrix or fax printing and handwriting, they can be practically illegible. The second problem we encountered was the structural complexity of the documents themselves. For example, just over 50 percent of the documents contained marginalia of some type, such as filing data, distribution lists, stamps of various types, editing, and handwritten comments. Most documents contained large amounts of peripheral data like names, dates, addresses, and distribution lists. Although these features are significant for archiving, they have little or no rhetorical value for our intended research. Other documents often contain or consist of forms, tables, and images that also offered little value for our analysis. To account for these problems we decided to keyboard the documents by hand as plain ASCII text and to code them with XML tags. When we investigated the existing XML tag sets, TEI in particular, we found that they are particularly well suited for archiving standard texts in standard hierarchies and for naming typesetting conventions. However, we chose to devise a set of tags specific to our project for primarily two reasons. First, we found that many of the documents collected for the corpus had a very non-standard format. In fact, we found no fixed definition for what constitutes a document in the tobacco archives. For example, a document recently coded began in the middle of a paragraph and sentence, proceeded for half of a page, changed to a summary of a court ruling, then to a policy letter, then to a table of denicotinization, and ended with a diagram of a processing facility. The second reason for devising our own tag set is that our primary interest is in archiving rhetorically significant text and events rather than the typesetting conventions used to represent them. Thus although TEI includes a full set of tags to indicate divisions and typographical conventions, use of these tags for our purposes might lead to ambiguities. For example, italics, boldface, and underlining have all been found to denote emphasis in the document set, which is rhetorically significant; however, they have also been found to denote titles, headings, names, quotations, formulas, and standard text, which may be of little value for our analysis. Thus, tagging a word in a document with a tag designed to denote typesetting, such as italics, may not be so useful when the corpus is analyzed linguistically or rhetorically simply because there is no way to know the significance of that particular event. To counter this, we have devised a set of XML tags which accommodates the structural complexity of the original documents and which reflects the purpose of our study. Data is retrieved from the XML files in a straightforward manner. We have embedded the expat XSLT engine into several Python scripts. This allows us to assemble a text corpus for study from the reference and manipulated-cases corpora according to the needs of the research. That is, with our scripts the XML files are parsed, desired tag content is selected, and the selected content is assembled and written to an ASCII text file. There are, however, two notable differences between the standard Web use of XSLT and that of our project. The first is data permanence. The output of our XSLT processing is ASCII/ANSI text which is written to file for later analysis rather than HTML sent onto the Internet. The other difference is that the XSLT output is not solely determined by the XSL stylesheet. For ease and speed in processing, some general document and tag selection is done by regular expressions in the Python script prior to calling the expat program. The end result of this initial phase of our project will be a larger general corpus of TDs for use as a reference, and a smaller corpus of “manipulated” TDs for focused analysis. Both will be archived as ASCII text with XML tags, which will allow us to generate tailored sub-corpora for specific studies using XSLT. Although these corpora are being created for our own purposes, our intent is to make them freely available to other researchers over the Internet. What we envision is an integrated Web site that provides access to the corpora in several formats: the TIF and/or PDF images of the original documents, the XML files coded with TEI compliant tags, the XML files coded with our tag set, ASCII text versions of the files, and access to a CGI version of our XSLT scripts for generating task-specific sub-corpora.