Treating texts as individuals vs. lumping them together

Ted Underwood has been talking up the advantages of the Mann-Whitney test over Dunning's Log-likelihood, which is currently more widely used. I'm having trouble getting M-W running on large numbers of texts as quickly as I'd like, but I'd say that his basic contention--that Dunning log-likelihood is frequently not the best method--is definitely true, and there's a lot to like about rank-ordering tests.

Before I say anything about the specifics, though, I want to make a more general point first, about how we think about comparing groups of texts.The most important difference between these two tests rests on a much bigger question about how to treat the two corpuses we want to compare.

Are they a single long text? Or are they a collection of shorter texts, which have common elements we wish to uncover? This is a central concern for anyone who wants to algorithmically look at texts: how far can we can ignore the traditional limits between texts and create what are, essentially, new documents to be analyzed? There are extremely strong reasons to think of texts in each of these ways.

Compare and Contrast

I may (or may not) be about to dash off a string of corpus-comparison posts to follow up the ones I've been making the last month. On the surface, I think, this comes across as less interesting than some other possible topics. So I want to explain why I think this matters now. This is not quite my long-promised topic-modeling post, but getting closer.

Off the top of my head, I think there are roughly three things that computers may let us do with text so much faster than was previously possible as to qualitatively change research.

1. Find texts that use words, phrases, or names we're interested in.
2. Compare individual texts or groups of texts against each other.
3. Classify and cluster texts or words. (Where 'classifying' is assigning texts to predefined groups like 'US History', and 'clustering' is letting the affinities be only between the works themselves).

These aren't, to be sure, completely different. I've argued before that in some cases, full-text search is best thought of as a way to create a new classification scheme and populating it with books. (Anytime I get fewer than 15 results for a historical subject in a ProQuest newspapers search, I read all of them--the ranking inside them isn't very important). Clustering algorithms are built around models of cross group comparisons; full text searches often have faceted group comparisons. And so on.

But as ideal types, these are different, and in very different places in the digital humanities right now. Everybody knows about number 1; I think there's little doubt that it continues to be the most important tool for most researchers, and rightly so. (It wasn't, so far as I know, helped along the way by digital humanists at all). More recently, there's a lot of attention to 3. Scott Weingart has a good summary/literature review on topic modeling and network analysis this week--I think his synopsis that "they’re powerful, widely applicable, easy to use, and difficult to understand — a dangerous combination" gets it just right, although I wish he'd bring the hammer down harder on the danger part. I've read a fair amount about topic models, implemented a few on text collections I've built, and I certainly see the appeal: but not necessarily the embrace. I've also done some work with classification.

In any case: I'm worried that in the excitement about clustering, we're not sufficiently understanding the element in between: comparisons. It's not as exciting a field as topic modeling or clustering: it doesn't produce much by way of interesting visualizations, and there's not the same density of research in computer science that humanists can piggyback on. At the same time, it's not nearly so mature a technology as search. There are a few production quality applications that include some forms of comparisons (WordHoard uses Dunning Log-Likelihood; I can only find relative ratios on the Tapor page). But there isn't widespread adoption, generally used methodologies for search, or anything else like that.

This is a problem, because cross-textual comparison is one of the basic competencies of the humanities, and it's one that computers ought to be able to help with. While we do talk historically about clusters and networks and spheres of discourse, I think comparisons are also closer to most traditional work; there's nothing quite so classically historiographical as tracing out the similarities and differences between Democratic and Whig campaign literature, Merovingian and Carolingian statecraft, 1960s and 1980s defenses of American capitalism. These are just what we teach in history---I in fact felt like I was coming up with exam or essay questions writing that last sentence.

So why isn't this a more vibrant area? (Admitting one reason might be: it is, and I just haven't done my research. In that case, I'd love to hear what I'm missing).

Dunning Amok

A few points following up my two posts on corpus comparison using Dunning Log-Likelihood last month. Nur ein stueck Technik.

Ted said in the comments that he's interested in literary diction.

I've actually been thinking about Dunnings lately too. I was put in mind of it by a great article a couple of months ago by Ben ZimmerZimmerman addressing the character of "literary diction" in a given period (i.e., Dunnings on a fiction corpus versus the broader corpus of works in the same period).

I'd like to incorporate a diachronic dimension to that analysis. In other words, first take a corpus of 18/19c fiction and compare it to other books published in the same period. Then, among the words that are generally overrepresented in 18/19c fiction, look for those whose degree of overrepresentation *peaks in a given period* of 10 or 20 years. Perhaps this would involve doing a kind of meta-Dunnings on the Dunnings results themselves.

I'm still thinking about this, as I come back to doing some other stuff with the Dunnings. This actually seems to me like a case where the Dunning's wouldn't be much good; so much of a Dunning score is about the sizes of the corpuses, so after an initial comparison to establish 'literary diction' (say), I think we'd just want to compare the percentages.

Theory First

Natalie Cecire recently started an important debate about the role of theory in the digital humanities. She's rightly concerned that the THATcamp motto--"more hack, less yack"--promotes precisely the wrong understanding of what digital methods offer:

the whole reason DH is theoretically consequential is that the use of technical methods and tools should be making us rethink the humanities.

Cecire wants a THATcamp theory, so that the teeming DHers can better describe the implications of all the work that's going on. Ted Underwood worries that claims for the primacy of theory can be nothing more than a power play, serving to reify existing class distinctions inside the academy; but he's willing to go along with a reciprocal relation between theory and practice going forward.

Dunning Statistics on authors

As promised, some quick thoughts broken off my post on Dunning Log-likelihood. There, I looked at _big_ corpuses--two history classes of about 20,000 books each. But I also wonder how we can use algorithmic comparison on a much smaller scale: particularly, at the level of individual authors or works. English dept. digital humanists tend to rely on small sets of well curated, TEI texts, but even the ugly wilds of machine OCR might be able to offer them some insights. (Sidenote--interesting post by Ted Underwood today on the mechanics of creating a middle group between these two poles).

As an example, let's compare all the books in my library by Charles Dickens and William Dean Howells, respectively. (I have a peculiar fascination with WDH, regular readers may notice: it's born out of a month-long fascination with Silas Lapham several years ago, and a complete inability to get more than 10 pages into anything else he's written.) We have about 150 books by each (they're among the most represented authors in the Open Library, which is why I choose it), which means lots of duplicate copies published in different years, perhaps some miscategorizations, certainly some OCR errors. Can Dunning scores act as a crutch to thinking even on such ugly data? Can they explain my Howells fixation?

I'll present the results in faux-wordle form as discussed last time. That means I use wordle.com graphics, but with the size corresponding not to frequency but to Dunning scores comparing the two corpuses. What does that look like?

Comparing Corpuses by Word Use

Historians often hope that digitized texts will enable better, faster comparisons of groups of texts. Now that at least the 1grams on Bookworm are running pretty smoothly, I want to start to lay the groundwork for using corpus comparisons to look at words in a big digital library. For the algorithmically minded: this post should act as a somewhat idiosyncratic approach to Dunning's Log-likelihood statistic. For the hermeneutically minded: this post should explain why you might need _any_ log-likelihood statistic.

What are some interesting, large corpuses to compare? A lot of what we'll be interested in historically are subtle differences between closely related sets, so a good start might be the two Library of Congress subject classifications called "History of the Americas," letters E and F. The Bookworm database has over 20,000 books from each group. What's the difference between the two? The full descriptions could tell us: but as a test case, it should be informative to use only the texts themselves to see the difference.

That leads a tricky question. Just what does it mean to compare usage frequencies across two corpuses? This is important, so let me take this quite slowly. (Feel free to skip down to Dunning if you just want the best answer I've got.) I'm comparing E and F: suppose I say my goal to answer this question:

What words appear the most times more in E than in F, and vice versa?

There's already an ambiguity here: what does "times more" mean? In plain English, this can mean two completely different things. Say E and F are exactly the same overall length (eg, each have 10,000 books of 100,000 words). Suppose further "presbygational" (to take a nice, rare, American history word) appears 6 times in E and 12 times in F. Do we want to say that it appears two times more (ie, use multiplication), or six more times (use addition)?

Bookworm and library search

We just launched a new website, Bookworm, from the Cultural Observatory. I might have a lot to say about it from different perspectives; but since it was submitted to the DPLA beta sprint, let's start with the way it helps you find library books.

Google Ngrams, which Bookworm in many ways resembles, was fundamentally about words and their histories; Bookworm tries to place texts much closer to the center instead. At their hearts, Ngrams uses a large collection of texts to reveal trends in the history of words; Bookworm lets you use words to discover the history of different groups of books--and by extension, their authors and readers.

Is catalog information really metadata?

We've been working on making a different type of browser using the Open Library books I've been working with to date, and it's raised a interesting question I want to think through here.

I think many people looking at word countson a large scale right now (myself included) have tended to make a distinction between wordcount data on the one hand, and catalog metadata on the other. (I know I have the phrase "catalog metadata" burned into my reflex vocabulary at this point--I've had to edit it out of this very post several times.) The idea is that we're looking at the history of words or phrases, and the information from library catalogs can help to split or supplement that. So for example, my big concern about the ngrams viewer when it came out was that it included only one form of metadata (publication year) to supplement the word-count data, when it should really have titles, subjects, and so on. But that still assumes that word data--catalog metadata is a useful binary.



I'm starting to think that it could instead be a fairly pernicious misunderstanding.

Wars, Recessions, and the size of the ngrams corpus

Hank wants me to post more, so here's a little problem I'm working on. I think it's a good example of how quantitative analysis can help to remind us of old problems, and possibly reveal new ones, with library collections.

My interest in texts as a historian is particularly focused on books in libraries. Used carefully, an academic library is sufficient to answer many important historical questions. (That statement might seem too obvious to utter, but it's not--the three most important legs of historical research are books, newspapers, and archives, and the archival leg has been lengthening for several decades in a way that tips historians farther into irrelevance.) A fair concern about studies of word frequency is that they can ignore the particular histories of library acquisition patterns--although I think Anita Guerrini takes that point a bit too far in her recent article on culturomics in Miller-McCune. (By the way, the Miller-McCune article on science PhDs is my favorite magazine article of the last couple of years). A corollary benefit, though, is that they help us to start understanding better just what is included in our libraries, both digital and brick.

Background: right now, I need a list of of the most common English words. (Basically to build a much larger version of the database I've been working with; making it is teaching me quite a bit of computer science but little history right now). I mean 'most common' expansively: earlier I found that about 200,000 words gets pretty much every word worth analyzing. There were some problems with the list I ended up producing. The obvious one, the one I'm trying to fix, is that words from the early 19th century, when many fewer books were published, will be artificially depressed compared to newer ones.

But it turns out that a secular increase in words published per year isn't the only effect worth fretting about. Words in the Google Books corpus doesn't just increase steadily over time. Looking at the data series on overall growth, one period immediately jumped out at me:

Graphing and smoothing

I mentioned earlier I've been rebuilding my database; I've also been talking to some of the people here at Harvard about various follow-up projects to ngrams. So this seems like a good moment to rethink a few pretty basic things about different ways of presenting historical language statistics. For extensive interactions, nothing is going to beat a database or direct access to text files in some form. But for quick interactions, which includes a lot of pattern searching and public outreach, we have some interesting choices about presentation.

This post is mostly playing with graph formats, as a way to think through a couple issues on my mind and put them to rest. I suspect this will be an uninteresting post for many people, but it's probably going to live on the front page for a little while given my schedule the next few weeks. Sorry, visitors!

Moving

Starting this month, I’m moving from New Jersey to do a fellowship at the Harvard Cultural Observatory. This should be a very interesting place to spend the next year, and I’m very grateful to JB Michel and Erez Lieberman Aiden for the opportunity to work on an ongoing and obviously ambitious digital humanities project. A few thoughts on the shift from Princeton to Cambridge:

What's new?

Let me get back into the blogging swing with a (too long—this is why I can't handle Twitter, folks) reflection on an offhand comment. Don't worry, there's some data stuff in the pipe, maybe including some long-delayed playing with topic models.

Even at the NEH's Digging into Data conference last weekend, one commenter brought out one of the standard criticisms of digital work—that it doesn't tell us anything we didn't know before. The context was some of Gregory Crane's work in describing shifting word use patterns in Latin over very long time spans (2000 years) at the Perseus Project: Cynthia Damon, from Penn, worried that "being able to represent this as a graph instead by traditional reading is not necessarily a major gain." That is to say, we already know this; having a chart restate the things any classicist could tell you is less than useful. I might have written down the quote wrong; it doesn't really matter, because this is a pretty standard response from humanists to computational work, and Damon didn't press the point as forcefully as others do. Outside the friendly confines of the digital humanities community, we have to deal with it all the time.

Predicting publication year and generational language shift

Before end-of-semester madness, I was looking at how shifts in vocabulary usage occur. In many cases, I found, vocabulary change doesn't happen evenly across across all authors. Instead, it can happen generationally; older people tend to use words at the rate that was common in their youth, and younger people anticipate future word patterns. An eighty-year-old in 1880 uses a world like "outside" more like a 40-year-old in 1840 than he does like a 40-year-old in 1880. The original post has a more detailed explanation.

Will had some some good questions in the comments about how different words fit these patterns. Looking at different types of words should help find some more ways that this sort of investigation is interesting, and show how different sorts of language vary. But to look at other sorts of words, I should be a little clearer about the kind of words I chose the first time through. If I can describe the usage pattern for a "word like 'outside'," just what kind of words are like 'outside'? Can we generalize the trend that they demonstrate?

The 1940 election

A couple weeks ago, I wrote about how ancestry.com structured census data for genealogy, not history, and how that limits what historians can do with it. Last week, I got an interesting e-mail from IPUMS, at the Minnesota population center on just that topic:

We have an extraordinary opportunity to partner with a leading genealogical firm to produce a microdata collection that will encompass the entire 1940 census of population of over 130 million cases. It is not feasible to digitize every variable that was collected in the 1940 census. We are therefore seeking your help to prioritize variables for inclusion in the 1940 census database.

In search of the great white whale

All the cool kids are talking about shortcomings in digitized text databases. I don't have anything so detailed to say as what Goose Commerce or Shane Landrum have gone into, but I do have one fun fact. Those guys describe ways that projects miss things we might think are important but that lie just outside the most mainstream interests—the neglected Early Republic in newspapers, letters to the editor in journals, etc. They raise the important point that digital resources are nowhere near as comprehensive as we sometimes think, which is a big caveat we all need to keep in mind. I want to point out that it's not just at the margins we're missing texts: omissions are also, maybe surprisingly, lurking right at the heart of the canon. Here's an example.