You are looking at archived content from my "Bookworm" blog, an experiment that ran from 2014-2016. Not all content may work. For current posts, see here.

Writing up text analysis for immediate interaction <em>and</em> long-term persistence.

Apr 19 2015

Though more and more outside groups are starting to adopt Bookworm for their own projects, I haven’t yet written quite as much as I’d like about how it should work. This blog is attempt to rectify that, and begin to explain how a combination of blogging software, interactive textual visualizations, and a exploratory data analysis API for bag-of-words models can make it possible to quickly and usefully share texts through a Bookworm installation.

But this is a difficult task. So much so that I have to completely change my blogging stack to do it. So for a first post on this site, I want to introduce some elements of the API and talk about why I think a platform like this is valuable for exploring a large collection of texts visually and quantitatively. Maybe someone will be persuaded to do it themselves.

This is going to be a somewhat technical post (I’ll try to keep the really technical stuff in the footnotes), but let me start by showing what we’re headed.

Note: even if you don’t care about the infrastructure and code that makes up much of this post, you may want to scroll down for a couple of the interactives: language use by popular directors, and a chart that shows the vocabulary used to describe liberal and conservative faculty members on their teaching evaluations.

I want to start thinking about how to reconcile two somewhat distinct traditions in Digital Humanities text-oriented projects.

One is computationally inclined, shies in the direction of writing articles, performs sometimes elaborate computations, and views reproducible research (where possible) as a matter of sharing code on Github. The other is more interested in presenting individual artifacts through robust platforms like Omeka for images or resilient sites for displaying TEI archives. The latter rarely tries to make broader arguments about its aggregate sources; the former rarely makes its underlying texts accessible except as a few examples in its text, relying instead largely on flat charts from R or python.

I count myself, and my two primary blogs, in the former camp. Sustainability and broad accessibility are awkward questions I’ve put off to deal with later as we rush to do amazing new things we can do with million-document collections. Recently I’ve been pushed more and more in the direction of web-based interactivity; but sustainability has remained off my horizon.

But the time has come to think through, a little more, how to keep some of these elements around for the long term. The big question this platform struggles with is much broader: how can a platform for textual visualization and criticism be:

  1. Richly interactive for years through access to terabytes of information on a non-cachable deep web backend, so long as the computers are plugged in and given minimal maintenance.
  2. Capable of persisting for decades even off the original server with its data backend, through locally cached copies on places like the Internet Archive and Github.
  3. As little reliant on the distributed architecture of the Internet era as possible. Large portions of the web would break if Google or JQuery started hosting their minified javascripts at a slightly different URL.

So those are the goals: the platform consists of three components. From stable to experimental, they are:

  1. Something established: the Bookworm platform for text analysis and visualization, currently used as a research and discovery platform by organizations like the Hathi Trust Research Center and the Yale University Libraries. This has its own stack you can read about in the documentation.
  2. Something less established but you may have seen: my D3-reimagining of the Bookworm front-end to enable many new forms of visualizations, like animated maps of the State of the Union or dotplots of differences in vocabulary used in course evaluations of male and female professors.
  3. Something entirely new: A hakyll-based blogging platform to easily share investigations on a Bookworm server. Source code is available on github. This may not be for everyone, though I hope some other places may want to adopt it. In other cases, rather than hakyll it might make sense to use a local CMS of any sort with embedded Bookworm elements. The basic principles will be the same.

Last fall, I released a bookworm for looking at trends in language use across 80,000 movies and TV episodes. That contained a single interactive chart without much text for exploring a vast database.

What I’m going to be talking about here is a different way for the people who want to maintain a Bookworm installation like that to explore their data. Instead of presenting just the data visualization as a blank canvas, I’m going to be talking about a way to write up what’s in that sort of database in a more narrative, but still interactive way. So for example: if I want to make a point about how much different directors use the word “love,” I can custom-write a query and present you with a barchart to look at it.

Relative usage of the word “love” in the films of several prolific directors

"director":["Allen, Woody","Ford, John (I)","Eastwood, Clint (I)","Huston, John (I)","Herzog, Werner (I)","Spielberg, Steven","Scorsese, Martin","Wilder, Billy","Preminger, Otto","Altman, Robert (I)","Hawks, Howard","Soderbergh, Steven","Nichols, Mike (I)","Burton, Tim (I)","Coppola, Francis Ford","Sirk, Douglas","Nolan, Christopher (I)","Bigelow, Katherine (I)"]

But unlike most inline graphics on a blog, this is powerfully interactive; you can click on any bar to see all of the underlying movies, for example.

Plus, since I’ve added a text box, you can search for anything you want. And a little javascript on my end adds more options. For instance: who do you think uses the word “death” the most? You can go back up to the text box and run an entirely different search: or just click one of the buttons below.

Change to ‘death’
Change to ‘money’
Change to ‘America’
Change to ‘United States’

This is a fun trick: but it’s also critically important for better critical engagement of research in the humanities. Facilitating this kind of open-ended interaction allows investigation on claims we make. Digital Humanists sometimes take to mean “reproducible research” as putting code on GitHub. That’s great and important; but it’s also a huge barrier to pass, and often an unnecessary one. You can learn a lot about a source by exposing different elements for research interactively, and that means that domain experts who will never in a million years run our LAMP-stack architecture can just put some queries in and see what the results are; they can also click through to the texts to see if the words mean the things I say they do.

The Bookworm Visualization API.

I’ve already described Bookworm’s core API elsewhere., but let me just briefly give a simple example. Let’s say that you want to know how many times each of the authors of the Federalist Papers uses the word “upon.” If you have a bookworm instance that includes all the Federalist Papers (which is easy: it’s the default test suite), you would express that this way: limit your search to the word “upon,” specify that you want responses grouped by the name of the author, and then ask for a counttype of “wordcount.”

    "database": "federalist",
    "search_limits": { "word": ["upon"] },
    "groups": ["author"],
    "counttype": ["WordCount"],
    "method": "json"

Deployed against a Bookworm endpoint, you’ll get the following results back:

			{ "DISPUTED": [3], "MADISON": [7], "HAMILTON": [374], "COAUTHORED": [2], "JAY": [1] }

That is: Hamilton uses the word “upon” 374 times: the other two authors use about 10 times at most.

Stacking Visualization on the Basic API

It’s ridiculous, nowadays, to display that sort of information as a table. While the API returns numbers, we also want visualization to be a first-class result. The Bookworm visualization API used here is essentially a lightweight set of additions to that core API to include a vocabulary for making a number of charts usefuly specifically for text analysis in the presence of metadata.

It borrows from the “graphics of graphics” described by Leland Wilkinson and popularized by Hadley Wickham, while also including all of the rich filtering, algorithmic, and search tools from Bookworm’s own API. It adds two new core elements (along with several chart-specific ones I’ll get to later):

  1. plotType, by which you specify one of several plots you wish to create, and
  2. aesthetic, in which you create an aesthetic mapping of variables.

The API format for a chart just substitutes in those two new fields.

Here is the API specification to request a barplot of that same data.

    "database": "federalist",
    "plotType": "barchart",
    "search_limits": {
        "word": ["upon"]
    "aesthetic": { "x": "WordCount", "y": "author" }

How do you embed such a chart? In any website where you have the Bookworm D3 library and associated libraries (most notably the core D3 library) loaded, you can just load up the query and deploy it against a Bookworm item.

Embedding responsive Bookworm visualization with simple javascript.

			// Define a query
var query = {
    database: 'federalist',
    plotType: 'barchart',
    search_limits: {
        word: ['upon']
    aesthetic: { x: 'WordCount', y: 'author' }

// Initialize the Bookworm with a query
var newBookworm = Bookworm(query);

// Create a D3 selection with an SVG
var svg ='body').append('svg');

// Stamp the SVG with the chart that you want.

//If you change the query object, you can refresh the plot with:

If you want to build a robust, powerful visualization and you know some javascript, this provides a relatively easy way to do so without even knowing D3 at all.

You can just dynamically manipulate the bookworm element’s query object to create dynamic charts that are fully updateable with things like the button above.

Embedded visualizations without javascript coding.

In some cases, even that will be too hard: the digital humanities tend to lean on python and R as their default languages, and javascript (particularly seen through D3’s worldview) can be difficult. I myself don’t want to have to fire up javascript just to make a simple chart. Since the API is so expressive, it’s easy to just write these visualizatons as special blocks inside markdown. The current website uses a hakyll site (the haskell jekyll equivalent) in which you can simply drop in your queries as markdown blocks.

Here’s what that block looks like as Markdown, with the height and width specified in advance:

An example Markdown block from this blog.

			```{.bookworm width=400 height=300}
{ "database": "federalist",
"plotType": "barchart",
"search_limits": {
"word": ["upon"]},
"aesthetic": {  "x": "WordCount",  "y": "author" }

And here’s the navigable element it automatically produces in the browser, giving access to both the raw API call (so that others can see the exact definition you’re working with), an interactive SVG which you can click to see and read individual documents (so that domain experts who are not coding experts can explore your data).

			    { "database": "federalist",
    "plotType": "barchart",
    "search_limits": {
    "word": ["upon"]},
    "aesthetic": {  "x": "WordCount",  "y": "author" }

One of the most important elements of a Bookworm is that it frequently needs to be interactive. You don’t just want one chart: you want the potential to create millions with clean transitions between them. That’s possible in this through the addition of filters to your markdown writeup.

Filters are HTML elements bound to the Bookworm view that turn it into an interactive element on the page. See this footnote for a fuller description.

The final element looks like this. If you type elements into the box, the bars will automatically update to reflect whatever word you’ve chosen.

			    { "database": "federalist",
    "plotType": "barchart",
    "search_limits": {
    "word": ["upon"]},
    "aesthetic": {  "x": "WordCount",  "y": "author" }

Bar charts are a classic, and useful, way to plot: but they are only one of the many visualizations built in to Bookworm. To keep it on the federalist papers, here’s what that would look like. (Remember, you can toggle over to “code” if you want to see the exact parameters of the call.

Relative use of Congress (x-axis) by Federalist Paper Number (y-axis)

			    { "database": "federalist",
    "plotType": "pointchart",
    "search_limits": {
    "word": ["Congress"]},
    "aesthetic": {  "x": "WordsPerMillion",  "y": "title" , "color":"author" }

Extremely different charts are easily specified by this. The most popular single browser so far probably has been my RateMyProfessor browser, using which uses paired dots to represent usage rates in reviews of male and female professors. The create call for this is relatively simple. Just click on “interactive SVG” to see the actual chart.

Rate My Professor: Language use by gender and department.

			    { "database": "RMP",
    "plotType": "pointchart",
    "search_limits": {"department__id":{"$lte":20},
    "word": ["funny"]},
    "aesthetic": {  "x": "WordsPerMillion",  "y": "department" , "color":"gender" }

But a powerful API means that there are many, many more ways to look at this. Just a line of difference makes this a comparison of Democrats and Republicans:

Rate My Professor: Descriptions by political party and department

			    { "database": "RMP",
    "plotType": "pointchart",
    "search_limits": {"department__id":{"$lte":20},"party":["D","R"],
    "word": ["old"]},
    "aesthetic": {  "x": "WordsPerMillion",  "y": "department" , "color":"party" }

A fully-featured interactive chart-creation engine, I’m convinced, would be a less-than-ideal outcome. (Not for lack of trying: I’ve had versions percolating for two or three years now.) It would be easy to add a dropdown to shift from one of these to the other; but ridiculous to allow a dropdown where the dots could be dozens of different schools. This chart kind of lets you know that Valencia college has a lot of online education: but it’s far from a good way to demonstrate it.

A terrible chart possible under the API: word usage in reviews at the top 20 schools.

			    { "database": "RMP",
    "plotType": "pointchart",
    "search_limits": {"department__id":{"$lte":20},"school__id":{"$lte":20},
    "word": ["online"]},
    "aesthetic": {  "x": "WordsPerMillion",  "y": "department" , "color":"school" }

A sufficiently powerful API will always allow this sort of junk to be created. I increasingly think that some human thought should go into deciding what a useful version of a visualization like this should be. Part of persuading people about conclusions on large data sets should be giving them tools to see if your interpretation holds up under scrutiny.


If you’re coming to this post from the world of data visualization, rather than Digital Humanities, you may notice that by having a number of separate svg items stack atop each other I’m violating the current spirit of the times, which prefers to have a single SVG that is updated inline in response to scroll events.

  1. For this form, narrative text comes first and foremost. Bookworm is great at showing charts; only text is any good at explaining the most interesting ones to the sort of people who can’t read graphs well, which includes many, many humanists.
  2. Pragmatically: I haven’t yet mastered the entire art of event listeners that will be necessary to make this work.
  3. Most importantly, though, I want a format that can at least gesture towards preservability. So that’s the rest of the post here.

These interactive graphics are served off a custom API using a running MySQL instance that can be terabytes large. There’s good reason to be hopeful that, when a Bookworm instance is useful, we can find ways to make it persist for years. But problems happen–when both Erez and I moved institutions in 2013 in a period when the Bookworm project was not receiving any immediate grant funding, it took about 2 or 3 months to restore the largest installation.

In addition to these full measures, I want to take half-measures as well as a failsafe plan. The current state of the Internet Archive cannot cache all the dark content on one of our servers. But it can cache a webpage, with narrative, with a bunch of static PNGs. This site uses a static page-generation framework to make it much easier to cache versions of the page. It also includes a number of scripts for automatically rendering PNGs of the state for every image in the text and embedding it directly in, so that at least the graphical element of the page can remain for as long as web archiving practices persists. These are far less powerful and useful than the interactive images: but at least they mean that a Bookworm page like this continues to meet the basic threshold for sustainability of, say, a pdf on even as it allows much greater interactivity.

PNGs aren’t the be-all and end-all. If anyone cares about the Internet in 30 years as much as we care about 80s Nintendo games now, there’s good reason to think that arbitrary HTML5/javascript will still run if the various off-site links still work. I’m assuming the underlying database is gone: but by caching the data from the calls saved explicitly specified, we can ensure that at least some of the most interesting statistics on any given blog post persist. It also means, after a little bit more work to the Bookworm-D3 plugin, that we’ll be able to .

In some cases, we may even be able to cache the entirety of every search an individual might run. The State of the Union Interactives Mitch Fraas and I built for the Atlantic adopt this strategy: Chris Barna figured out a great system where the front-end Bookworm infrastructure is hosted on the Atlantic site, and he simply cached every possible query including search results that might come through the Atlantic’s servers and pulled it from my local State of the Union server. As a result, that Atlantic site now has no Bookworm backend at all, though it successfully uses the API. As a general strategy, this will work only when there aren’t fantastically rich calls being generated. (Anything with an open-entry text box is unlikely to work.)

But even for open-text Bookworms like my Rate My Professor browser, there’s some hope for a useful long-term persistence. Thanks to query caching, I now have a list of the tens of thousands of most popular search terms for faculty members. Even if I choose to retire the completely interactive portion of the web site, I can still keep a static one running that will include almost all the entries that actually human beings tend to search for.

In essence, this mimics a sort of process of peer review–there is open interaction and comment for the first few years of a project, but then it gets locked into a more fixed form for the long term unless someone (the author, the press, the library) keeps investing energy to keep it up-to-date. The question of who will store it remains: but while it’s easy to get all Mad Max about the dismal future of digital sources, I tend to be on the more optimistic side. But maybe I’m wrong!