Some long term perspective on the "crisis" in humanities enrollment

There was an article in the Wall Street Journal about low enrollments in the humanities yesterday. The heart of the story is that the humanities resemble the late Roman Empire, teetering on a collapse precipitated by their inability to get jobs like those computer scientists can provide. (Never mind that the news hook is a Harvard report about declining enrollments in the humanities, which makes pretty clear that the real problem is students who are drawn to social sciences, not competitition from computer scientists.)

But to really sell a crisis, you need some numbers. Accompanying this was a graph credited to the American Academy of Arts and Sciences showing a spectacular collapse in humanities enrollments. I happen to have made one of the first versions of this chart working there several years ago. Although it shows up in the press periodically to enforce a story of decay, some broader perspective on the data makes clear that the "Humanities in crisis" story has the wrong interpretation, the wrong baseline, and the wrong denominator.


The interpretation: The chart never quite reinforces the point that something terrible is going on in the humanities right now. Anyone looking at it closely will notice, as Michael Bérubé has, that the real collapse of humanities enrollments happened in the 1970s. The Great Recession has been less ruinous to enrollments than were the mid-1990s. Sure, a few Harvard majors have switched from history to government in the last decade: is that really a story?

But it does succeed in making the humanities appear massively out of date. And that's a compelling story for all sorts of people. It makes humanists feel as though they deserve a larger share of the university, and that their sense of being under seige is due to the some pathology in the culture at large; it makes traditionalist critics of the humanities feel secure in pointing out that something has gone very wrong in the field.


As for the baseline: the chart starts in 1967 because that's the farthest back point that the federal government has online computerized records of enrollment. But baselines matter, a lot. This graph is misleading as a statement about what the humanities should look like in American life, or even what they used to look like.
1967 is not when they started tracking enrollment numbers--it's just the point where it's extremely easy to make a chart without leaving your computer. A long time ago, I went through two decades of paper printouts from the Bureau of Education and copied by hand what enrollments looked like before 1967. If any period is a very strange abberation in the history of the American university, it's the 1970s, not the 2010s. (It should be clear, but these are stacked--about 10% of all degrees were in the big humanities fields in the late 1940s).

The WSJ chart starts at precisely the moment of highest enrollment.

Degrees are, a bit, lower as a percentage of the university than they were in 1948: humanities were about 10% back then, and more like 8% in recent years. (These four majors form the vast majority of all humanities degrees--we're not losing much quantity by omitting music, Af-Am, and the rest. And I should note I'm speculating a bit about the total BA numbers before 1955 or so--they kept some first professional degrees lumped in, so I might be off by a percentage point or two. That's why this data couldn't be included in the official version of the Indicators, if I recall correctly).

And "percentage of all degrees" is a strange denominator. Everyone knows that they quantity and type of American universities has been changing dramatically. Compared to the massive changes in the American university since the Second World War, it's the resilience of the humanities that should be surprising. If what you care about is humanistic education, you shouldn't be worrying about market share inside the university, which is a dramatically changing thing. You should care about the whole population. And while the 60s boom is still larger than anything we see today, that shows the humanities stabilizing in a fairly healthy place through the early 2000s. (I don't have the data after that on hand, but to repeat--there is no non-anecdotal evidence of a steep drop in humanities enrollments in the last 10 years). To see the data from 2002 to 2011, read my revised version of this post at Edge of the American West.

A decline isn't a crisis.

This chart also presents some big problems for all the pat stories we want to tell about the humanities. Look at the period of their great recovery: it's the late 1980s and early 1990s--in other words, the heart of the culture wars, perhaps the only period that everyone agrees was ruinous to the humanities. Should we conclude we need to get back to some deconstruction in our courses? That seems unlikely.

We shouldn't be assessing the health of the humanities by market-share metrics that are far more about demographics and the changing face of higher ed than they are about the intellectual shifts at the heart of actual humanities practice.

Talking about crisis doesn't help us. Humanists obviously get great pleasure from describing themselves on the knife's edge, but they're not especially effective at mobilizing the language of crisis to actually advance our fields.

I suspect we're probably better if we don't buy into the narrative of collapse at all. More people are majoring in humanities fields. More books are being published in them. Whatever problems we have, they're not really about quantity. A fixation on corporatist measures of market share as representing the success of these fields is completely contrary to their aspirations. New enrollments in graduate programs are almost certainly not materially helping the field, but we too often act as if they are.
 
It's a non-story to say that the humanities are a minor but important part of the economy of the changing university. Everyone loves a narrative. But when someone tells you the humanities are collapsing, it's worth remembering that even the universities, not to even mention the country as a whole, have never been the bastions of humanistic learning we nostalgically invent.

Turning-point years in history

What are the major turning points in history? One way to think about that is to simply look at the most frequent dates used to start or end dissertation periods.* That gives a good sense of the general shape of time.

*For a bit more about how that works, see my post on the years covered by history dissertations: I should note I'm using a better metric now that correctly gets the end year out of text strings like "1848-61."

Here's what that list looks like: the most common year used in dissertation titles. It's extremely spiky--some years are a lot more common than are others.

What years do historians write about?

Here's some inside baseball: the trends in periodization in history dissertations since the beginning of the American historical profession. A few months ago, Rob Townsend, who until recently kept everyone extremely well informed about professional trends at American Historical Association* sent me the list of all dissertation titles in history the American Historical Association knows about from the last 120 years. (It's incomplete in some interesting ways, but that's a topic for another day). It's textual data. But sometimes the most interesting textual data to analyze quantitatively are the numbers that show up. Using a Bookworm database, I just pulled out from the titles the any years mentioned: that lets us what periods of the past historians have been the most interested in, and what sort of periods they've described..

*Townsend is now moving on to the American Academy of Arts and Sciences, where I'm excited to see that he'll manage the Humanities Indicators—my first real programming/data project was putting together the first version of them together with Malcolm Richardson immediately after college.

Numbers between 500 and 2000 are almost always years. You can see here that the vast bulk of historical study has been in the period since 1750: the three spikes out of the landscape correspond to the Civil War and the two world wars. Output decreases in the late 20th century in large part because the data set goes back to about 1850; but as we'll see in the next chart, not entirely.

How not to topic model: an introduction for humanists.

The new issue of the Journal of Digital Humanities is up as of yesterday: it includes an article of mine, "Words Alone," on the limits of topic modeling. In true JDH practice, it draws on my two previous posts on topic modeling, here and here. If you haven't read those, the JDH article is now the place to go. (Unless you love reading prose chock full've contractions and typos. Then you can stay in the originals.) If you have read them, you might want to know what's new or why I asked the JDH editors to let me push those two articles together. In the end, the changes ended up being pretty substantial.

Patchwork Libraries

The hardest thing about studying massive quantities of digital texts is knowing just what texts you have. This is knowlege that we haven't been particularly good at collecting, or at valuing.


The post I wrote two years ago about the Google Ngram chart for 02138 (the zip code for Cambridge, MA) remains a touchstone for me because it shows the ways that materiality, copyright, and institutional decisions can produce data artifacts that are at once inscrutable and completely understandable. (Here's the chart--go read the original post for the explanation.)



Since then, I've talked a lot about the need to understand both the classification schemes for individual libraries, and the need to understand the complicated historical provenance of the digital sources we use.

What I haven't done is give a real account of the sources of the books in Bookworm. That's pretty hypocritical. I apologize. I decided some time ago to use the Open Library to provide all the metadata for Bookworm. It has very good, high-quality library metadata; but it doesn't indicate where the metadata, or the volume itself, came from. (The meta-metadata, if you will, is not great.) That's an ontological issue. Open Library describes "editions," so there's no space for a field that tells you where an individual pdf or text volume came from.

I recently loaded in this library data, though. The Internet Archive web site, which stores the actual books, does say where a particular volume comes from. So with a little behind the scenes magic, it's pretty easy to get that into the database in some form.

I've been learning some D3 mock up a possible multi-dimensional interfaces to the Bookworm data through the API. (If you come to the Bookworm booth at the DPLA meeting next month, you can play with this as well as some new projects others in the lab have been building on the platform).

Using that, it's possible to quickly mock up a chart of the... (drumroll)...

 Library origins of the books in Bookworm. (Most common libraries only)

Click to enlarge

The colors are on a log scale here, and each little line represents a single combination of a library and year. So a red horizontal band is an area where a library has contributed hundreds of books to the Internet Archive (and therefore Bookworm) each year, a yellow to orange band means dozens of books a year, and the areas of scattered green show libraries that only contribute a volume or two every few years. I only show the most common libraries.

What can we see here? The number one contributor is the University of California Libraries. (No surprise). The  Robarts library at the University of Toronto is number two--that's only surprising to me because there have been so few books published in Canada when I've segmented by country. Most of the other large libraries are the ones you would expect—both some of the big university/Google partner libraries (Harvard, Oxford, Michigan, NYPL) and a few free agents who scanned on their own or in cooperation with the Internet Archive (Library of Congress, Boston Public (I think)).

The temporal patterns are more interesting. The 1922 cutoff line is the dominant visual feature of the chart--the extremely different composition of the lending libraries after that date (MIT and Montana State suddenly become important players, while sources like Cal and the LOC vanish). This is why comparisons across 1922 are never safe, even when the numbers seem big enough. But the complicated nature of the copyright line and scanning policies is also clear: for Canadiana.org, it's 1900, and for Oxford, it seems to be in the 1880s. There are clearly some metadata problems, as well (I think a lot of those post-1922 Harvard books are misattributed).

There are also a lot of strange little clusters of intense collection. Duke and the Lincoln Financial collection have massive spikes from about 1861-1865; the Boston Public Library seems to have an abnormal number of books right around 1620 and 1776. Those are specifically historical collections that will affect the sort of conclusions one can draw from aggregate data, much like the questions I was asking about a Google presentation that purported to show that more history was published in revolutionary times). Although oddly enough, the first few BPL volumes from 1620 I checked were in French or Latin--it's not quite the model of founding charters that I expected.

The digital libraries we're building are immense patchworks. That means they have seams. It's possible to tell stories that rest lightly across the many elements. But push too hard in the wrong direction, and conclusions can unravel. I don't think that any conclusions I've made in the past are unsettled by knowing where the books came from--although that Oxford dropoff has me wondering what might happen to transatlantic comparisons. But I'm glad to be able to see it, and so wanted to share. This is something we need to do.

And this isn't an issue particular to just Bookworm, although it does have a slightly more tangled line of transmission than some other libraries. Here, for example, are the top 16 libraries in the Hathi trust catalog from 1780 to 1820, scraped from their web catalog. 1800 is not a year that dramatically changes the publication history of books; but it is a year that librarians will use for their own arbitrary purposes. For instance, a book published in 1799 is considerably more likely to have been "rare" and therefore off limits to the Google scanners in the period of book scanning, I'd bet anything, just because human beings use rules of thumb like that; or in the period that university libraries were being built, acquisition policies may have been quite different for pre-1800 vs post-1800 texts in various genres. ("Don't buy any science books written before 1800" would be a sensible policy to adopt in 1870, for example. Almost all the libraries here would not have bought the bulk of their 1799 collections in the year 1799.)

Click to enlarge

I pulled out the year 1800 (1800 and 1900 are both massive spikes where all sorts of unknown books can be filed). That also serves to highlight the gaps from 1799 to 1801; they're sometimes quite significant. What you see here are the same sort of seams and discontinuities as in Bookworm, albeit on a smaller scale. The number of volumes from NYPL, Cal, and Harvard double overnight; other libraries, like Michigan and Madrid, seem not to show any pattern. If Michigan and California collect different sorts of books, this can cause major headaches for comparisons across the line.

There's no grand lesson here. Or if there is, it's just the old one: know your sources. But as a public service announcement, it's worth making again and again.

Canonic authors and the pronouns that they used

My last post had the aggregate statistics about which parts of the library have more female characters. (Relatively). But in some ways, it's more interesting to think about the ratio of male and female pronouns in terms of authors whom we already know. So I thought I'd look for the ratios of gendered pronouns in the most-collected authors of the late 19th and early twentieth centuries, to see what comes out.

On the one hand, I don't want to claim too much for this: anyone can go to a library and see that Washington Irving doesn't write female characters. But as one of many possible exercises in reducing down the size of the library to rethink the broad aspects of the literary canon, c. 1910, I do think it's suggestive; and, as I'll suggest towards the end, knowing these practical details can help us explore the instability of 'subject' or 'genre' as expressed by the librarians who choose where to put these books on the shelves.

Genders and Genres: tracking pronouns

Now back to some texts for a bit. Last spring, I posted a few times about the possibilities for reading genders in large collections of books. I didn't follow up because I have some concerns about just what to do with this sort of pronoun data. But after talking about it to Ryan Cordell's class at Northeastern last week, I wanted to think a little bit more about the representation of male and female subjects in late-19th century texts. Further spurs were Matt Jockers recently posted the pronoun usage in his corpus of novels; Jeana Jorgensen pointed to recent research by Kathleen Ragan that suggests that editorial and teller effects have a massive effect on the gender of protagonists in folk tales. Bookworm gives a great platform for looking at this sort of question.

Anachronism patterns suggest there's nothing special about words

I'm cross-posting here a piece from my language anachronisms blog, Prochronisms.

It won't appear the language blog for a week or two, to keep the posting schedule there more regular. But I wanted to put it here now, because it ties directly into the conversation in my last post about whether words are the atomic units of languages. The presumption of some physics inflected linguistics research is that it is. I was putting forward the claim that it's actually Ngrams of any length. This question is closely tied to the definition of what a 'word' is (although as I said in the comments, I think statistical regularities tend to happen at a level that no one would ever call a 'word,' however broad a definition they take).

The piece from Prochronisms is about whether writers have a harder time avoiding anachronisms when they appear as parts of multi-word phrases. Anachronisms are a great test case for observing what writers know about language. Writers are trying to talk as if they're from the past; but--and this is fundamental point I've been making over there--it's all but impossible for them to succeed. So by looking at failures, we can see at what level writers "know" language. If there's something special about words, we might expect them to know more about words than phrases. But if--as these preliminary data seem to indicate--users of language don't seem to have any special knowledge of individual words, that calls into question cognitive accounts of changes in language, like the one the physicists offered, that rely on some fixed 'vocabulary' limit that is enumerated in unigrams.

Anyhow, here's the Prochronisms post:

Are words the atomic unit of a dynamic system?

My last post was about how the frustrating imprecisions of language drive humanists towards using statistical aggregates instead of words: this one is about how they drive scientists to treat words as fundamental units even when their own models suggest they should be using something more abstract.

I've been thinking about a recent article by Alexander M. Petersen et al., "Languages cool as they expand: Allometric scaling and the decreasing need for new words." The paper uses Ngrams data to establish the dynamics for the entry of new words into natural languages. Mark Liberman argues that the bulk of change in the Ngrams corpus involves things like proper names and alphanumeric strings, rather than actual vocabulary change, which keeps the paper from being more than 'though-provoking.' Liberman's fundamental objection is that although the authors say they are talking about 'words,' it would be better for them to describe their findings in terms of 'tokens.' Words seem good and basic, but dissolve on close inspection into a forest of inflected forms, capitals, and OCR mis-readings. So it's hard to know whether the conclusions really apply to 'words' even if they do to 'tokens.'

Crossroads

Just a quick post to point readers of this blog to my new Atlantic article on anachronisms in Kushner/Spielberg's Lincoln; and to direct Atlantic readers interested in more anachronisms over to my other blog, Prochronisms, which is currently churning on through the new season of Downton Abbey. (And to stick around here; my advanced market research shows you might like some of the posts about mapping historical shipping routes.)

Keeping the words in Topic Models

Following up on my previous topic modeling post, I want to talk about one thing humanists actually do with topic models once they build them, most of the time: chart the topics over time. Since I think that, although Topic Modeling can be very useful, there's too little skepticism about the technique, I'm venturing to provide it (even with, I'm sure, a gross misunderstanding or two). More generally, the sort of mistakes temporal changes cause should call into question the complacency with which humanists tend to  'topics' in topic modeling as stable abstractions, and argue for a much greater attention to the granular words that make up a topic model.

In the middle of this, I will briefly veer into some odd reflections about how the post-lapsarian state of language. Some people will want to skip that; maybe some others will want to skip to it.

Military History and data: the US Navy in World War II

A stray idea left over from my whaling series: just how much should digital humanists be flocking to military history? Obviously the field is there a bit already: the Digital Scholarship lab at Richmond in particular has a number of interesting Civil War projects, and the Valley of the Shadow is one of the archetypal digital history projects. But it's possible someone could get a lot of mileage out of doing a lot more.

There are two opportunistic reasons to think so.

1. Digital historians have always been very interested in public audiences; military history has always been one of the keenest areas of public interest.

2. The data is there for algorithmic exploration. In most countries, no organization is better at keeping structured records than the military.

And the stuff is interesting. It's easy, for example,to pull out the locations of nearly the entire US Navy, season-by-season, in the Pacific Theater:

Click to enlarge.

Or even animate them and the less comprehensive Japanese records to show the tide of battle (America in blue, Japan in red):

Reading digital sources: a case study in ship's logs

Digitization makes the most traditional forms of humanistic scholarship more necessary, not less. But the differences mean that we need to reinvent, not reaffirm, the way that historians do history.


This month, I've posted several different essays about ship's logs. These all grew out of a single post; so I want to wrap up the series with an introduction to the full set. The motivation for the series is that a medium-sized data set like Maury's 19th century logs (with 'merely' millions of points) lets us think through in microcosm the general problems of reading historical data. So I want in this post to walk through the various parts I've posted to date as a single essay in how we can use digital data for historical analysis.

The central conclusion is this: To do humanistic readings of digital data, we cannot rely on either traditional humanistic competency or technical expertise from the sciences. This presents a challenge for the execution of research projects on digital sources: research-center driven models for digital humanistic resource, which are not uncommon, presume that traditional humanists can bring their interpretive skills to bear on sources presented by others.

All voyages from the ICOADS US Maury collection. Ships tracks in black, plotted on a white background, show the outlines of the continents and the predominant tracks on the trade winds. 

We need to rejuvenate three traditional practices: first, a source criticism that explains what's in the data; second, a hermeneutics that lets us read data into a meaningful form; and third, situated argumentation that ties the data in to live questions in their field.

Where are the individuals in data-driven narratives?

Note: this post is part 5 of my series on whaling logs and digital history. For the full overview, click here.

In the central post in my whaling series, I argued data presentation offers historians an appealing avenue for historical argumentation, analogous in importance to the practice of shaping personal stories into narratives in more traditional histories. Both narratives and data presentations can appeal to a broader public than more technical parts of history like historiography; and both can be crucial in making arguments persuasive, although they rarely constitute an argument in themselves. But while narratives about people ensure that histories are fundamentally about individuals, working with data generally means we'll be dealing with aggregates of some sort. (In my case, 'voyages' by 'whaling ships'.*)

*I put those in quotation marks because, as described at greater length in the technical methodology post, what I give are only the best approximations I could get of the real categories of oceangoing voyages and of whaling ships.

This is, depending on how you look at it, either a problem or an opportunity. So I want to wrap into this longer series a slightly abtruse--technical from the social theory side rather than the algorithmic side--justification for why we might not want to linger over individual experiences.

One major reason to embrace digital history is precisely that it lets us tell stories that are fundamentally about collective actions--the 'swarm' of the whaling industry as a whole--rather than traditional subjective accounts. While it's discomforting to tell histories without individuals, that discomfort is productive for the field; we need a way to tell those histories, and we need reminders they exist. In fact, those are just the stories that historians are becoming worse and worse at telling, even as our position in society makes us need them more and more.

When you have a MALLET, everything looks like a nail

Note: this post is part 4, section 2 of my series on whaling logs and digital history. For the full overview, click here.

One reason I'm interested in ship logs is that they give some distance to think about problems in reading digital texts. That's particularly true for machine learning techniques. In my last post, an appendix to the long whaling post, I talked about using K-means clustering and k-nearest neighbor methods to classify whaling voyages. But digital humanists working with texts hardly ever use k-means clustering; instead, they gravitate towards a more sophisticated form of clustering called topic modeling, particularly David Blei's LDA (so much so that I'm going to use 'LDA' and 'topic modeling' synonymously here). There's a whole genre of introductory posts out there encouraging humanists to try LDA: Scott Weingart's wraps a lot of them together, and Miriam Posner's is freshest off the presses.

So as an appendix to that appendix, I want to use ship's data to think about how we use LDA. I've wondered for a while why there's such a rush to make topic modeling into the machine learning tool for historians and literature scholars. It's probably true that if you only apply one algorithm to your texts, it should be LDA. But most humanists are better off applying zero clusterings, and most of the remainder should be applying several. I haven't mastered the arcana of various flavors of topic modeling to my own satisfaction, and don't feel qualified to deliver a full-on jeremiad against its uses and abuses. Suffice it to say, my basic concerns are:

1. The ease of use for LDA with basic settings means humanists are too likely to take its results as 'magic', rather than interpreting it as the output of one clustering technique among many.
2. The primary way of evaluating its result (confirming that the top words and texts in each topic 'make sense') ignores most of the model output and doesn't map perfectly onto the expectations we have for the topics. (A Gary King study, for example, that empirically ranks document clusterings based on human interpretation of 'informativeness' found Direchlet-prior based clustering the least effective of several methods.)

Ship data gives an interesting perspective on these problems. So, at the risk of descending into self-parody, I ran a couple topic models on the points in the ship's logs as a way of thinking through how that clustering works. (For those who only know LDA as a text-classification system, this isn't as loony as it sounds; in computer science, the algorithm gets thrown at all sorts of unrelated data, from images to music).

Instead of using a vocabulary of words, we can just use one of latitude-longitude points at decimal resolution. Each voyage is a text, and each day it spends in, say, Boston is one use of the word "42.4,-72.1". That gives us a vocabulary of 600,000 or so 'words' across 11,000 'texts', not far off a typical topic model (although the 'texts' are short, averaging maybe 30-50 words). Unlike k-means clustering, a topic model will divide each route up among several topics, so instead of showing paths, we can visually only look at which points fall into which 'topic'; but a single point isn't restricted to a single topic, so New York could be part of both a hypothetical 'European trade' and 'California trade' topic.

With words, it's impossible to meaningfully convey all the data in a topic model's output. Geodata has the nice feature that we can inspect all the results in a topic by simply plotting them on a map. Essentially, 'meaning' for points can be firmly reduced a two-dimensional space (although it has other ones as well), while linguistic meaning can't.

Here's the output of a model, plotted with high transparency so that a point on the map will appear black if it appears in that topic in 100 or more log entries. (The basic code to build the model and plot the code is here--dataset available on request).

Click to enlarge