On May 30, 2023, at 10:46 AM, Amy Schuler <000000088c12581f-dmarc-requ...@lists.clir.org> wrote:
> I'm interested in mining my email for phrases and concepts related to a > specific service that I've provided the past several years and for which I > have not kept good statistics (my bad). Being that library & information > services at my organization is a one-person shop (me), text mining my own > email may be helpful. I never delete email, so it's bound to be a rich and > messy corpus. I understand I may download my email archive using Google > Takeout, then I was thinking of using some R package like: > > https://github.com/matthewjdenny/REmail > > Does anyone have better, easier suggestions for tools to use, or experience > they would like to share? > > -- > Amy C. Schuler (she/her) > Director, Information Services & Library > > Cary Institute of Ecosystem Studies | 2801 Sharon Turnpike | Millbrook, NY > www.caryinstitute.org I recently finished writing a LibGuide on text mining and natural language processing. See: https://libguides.library.nd.edu/text-mining-and-nlp That said, I have a hammer and to me, everything looks like a nail. See also: https://reader-toolbox.readthedocs.io When doing text mining, the first thing one needs to do is ask themselves a research question they hope to address through text mining. The question can be as rudimentary as to determine the size of a collection, to denoting what a collection is about, to extracting a definition of social justice. What do you want to know? Second, you must point to content you believe can address the question. A set of books? A set of articles? A set of social media content? Apparently, the answer here is "my email". Third, you must get the content. Believe it or not, getting a set of digitized articles or books is more difficult than you might think. Go ahead. Do the perfect search against JSTOR and then download all the articles. Tedious, at best. The process is similar even when it comes to Project Gutenberg. In this case, you might already have the content, but you MUST convert it to plain text. When you look at raw email messages, they're a whole lot of noise. You will probably want to simply extract author, title, date, and content. Not trivial. Once you get this far, you will use any number of different techniques to model the text. They range from the simple to the complex: * counting & tabulating of unigrams (words) and bigrams (two-word phrases) * counting & tabulating of parts-of-speech (nouns, verbs, adjectives, etc) * counting & tabulating of named-entities (real-world things like people, places organizations, etc) * concordancing (think ^f on steroids) * extracting latent themes (topic modeling) and then pivoting the results over time, place, authors, etc * full-text searching complete with fields, Boolean logic, relevancy ranking, etc * semantic indexing (plotting words in an n-dimensional space and calculating which words are "near" other words * extracting sentences from the text, articulating grammars, and identifying sentences that match the grammars such as definitions, modalities, etc. * calculating statistically significant keywords (using something similar to TF/IDF), mapping those words to things in Wikidata, and discovering additional relationship (meanings) through Linked Data * identify words of interest, applying them to dictionaries or thesauri (like WordNet), to again, discovering additional relationships * repeat; this is an iterative process After applying one or more of the techniques above, the student, researcher, or scholar ought to be able to address their research question, but remember, like any modeling process, the results do not necessarily denote truth. Instead, the results provide one with observations, and the observations must be placed into context though interpretation. Simple solution: Once you get plain text, feed it to Voyant Tools, and you will probably go a long way to addressing your question(s). [1] I have done many different things in my career, and text mining has proven to be one of the more satisfying. Through text mining and natural language processing I have been able to read huge corpora (like eight different editions of Encyclopedia Britannica, the whole of Chinese history, or 100 years of United States Presidential papers). It is fun to compare and contrast ideas as they ebb and flow over time. It is fun to read an entire literature and make up my own mind what it means. For example, I assert Moby Dick is as much an instruction manual on whaling as it is about a man's relentless pursuit of the leviathan which chewed off his leg. Okay, my reply has probably been overkill, but I couldn't help myself. :) [1] Voyant Tool - https://voyant-tools.org/ -- Eric Lease Morgan Navari Family Center for Digital Scholarship Hesburgh Libraries University of Notre Dame
