They say the humanities are in "crisis." Society values the sciences more highly. Students turn to other majors promising apparently predictable paths to jobs. In recessions, universities and colleges cut humanities programs first. And funding support for the humanities continues to be a rounding error in the national budget. Decisively, they say, the humanities are in crisis. (See WE1S bibliography of "humanities crisis.")

So what does everyone actually say about the humanities? With what nuance, in what contexts, and in the company of what other, even contrasting views? And who are "they" anyway?

Our WhatEvery1Says (WE1S) project studied these questions by taking an approach complementing, but different from, surveying (though we did some surveying too). We read the media.In particular, we used computational methods to help us study vast numbers of articles and social media posts mentioning the humanities.

What everyone says about the humanities or anything else in the media, after all, is different from what anyone says one-on-one. The media is a reverberating echo chamber producing storylines through broadly impactful, but complicated, rhetorical scenarios—such as when a politician interviewed for a story addresses the generic taxpayer (or parent) in hopes of reaching a partisan voter. The apparently common wisdom that results from such rhetorical tangles is a media wisdom, a repertoire of memes that take on a life of their own and are hard to change because the myriad push-pull relations among all the stakeholders often lock winning memes in place. (See WE1S bibliography of "Journalism & Media")

Funded by $1.1 million from the Andrew W. Mellon Foundation, we gathered over 1 million (1,028,629) unique English-language journalistic media articles and related documents mentioning the literal word "humanities" (and for some research purposes also the "liberal arts," "the arts," and "sciences") from 1,053 U.S. and 437 international news and other sources from the 1980s through 2019, mostly after 2000 when news media began producing digital texts en masse. For comparison and other uses, we also gathered about 1.38 million unique documents representing a random sampling of all news articles. In addition, we harvested over 6 million posts mentioning the "humanities" and related terms from social media (about 5 million from Twitter and 1 million from Reddit); and about 1.2 million transcripts of U.S. television news broadcasts from those available in the Internet Archive. (See our Research Materials.)We focused on the U.S. to start with a manageable scope of materials, and also because our researchers—based at the partner institutions of U. California, Santa Barbara; California State U., Northridge; and U. Miami—have a poorer understanding of the social, economic, political, and cultural contexts of the media and the humanities in other nations. (We invite others to use our materials, methods, and tools and software to extend our research internationally.)

Why search for documents containing the literal word "humanities" and related words? We did so not because we think these terms by themselves cast a net over all the humanities. In the vast sea of public discourse, the humanities also appear under the names of "literature," "history," or other specific kinds; and the humanities are name-dropped everywhere in discussions of particular people, books, intellectual movements, controversies, organizations, events, and so on. The humanities are also surprisingly pervasive in media coverage of common life and death events—for example, in wedding announcements and the obituaries of everyday (not just famous) people. There is no predefined, bounded set of media documents for studying public discussion of the humanities.

So we aimed for a strategically chosen subset of materials mentioning the literal word "humanities" (plus sometimes other terms) on the hypothesis that we could thus capture a good swathe of examples on both sides of the line between a general concept of the humanities and specific kinds of humanities, and between wider public discussion of the humanities (as when "humanities" comes up in relation to words like "literature," "history," and people, books, organizations, etc.) and more specialized academic discourse on the humanities (as when universities call a group of disciplines "the humanities").

Then we sliced-and-diced our corpus (whole body of materials) into over thirty “collections,” each of which groups documents in different ways to facilitate asking different research questions. Our core study collection (Collection 1) consists of about 83,000 articles and other texts mentioning the “humanities” from U.S. mainstream, local, and higher-education student newspapers between c. 1989 and 2019. By contrast, for example, Collection 4 includes articles only from 15 top-circulation U.S. newspapers; Collection 14 restricts materials to U.S. college or university student newspapers; Collection 21 includes articles from top-circulation U.S. newspapers mentioning either “humanities” or “science”; Collection 29 includes Twitter posts between 2014 and 2017 containing "humanities" (aggregated by author); and Collection 38 includes Reddits longer than 225 words from 2006 to 2019 containing "humanities," "liberal arts," or "the arts." (See the metadata tags we also add to our data to help analyze groups of publication sources.)We feature nineteen of these collections on our Collections & Topic Models page, where we describe each and make them available for other researchers in the form of their word frequency counts, computational models and visualizations, and other derived or generated data (but, unfortunately, no readable text due to copyright or licensing constraints except for texts we directly gathered from public and open sources).

So that other researchers can fully benefit from the data we harvested and processed from the materials we studied, we deposited our full datasets in the Zenodo repository operated by CERN to promote "open science." These datasets (six in total) hold all the data and metadata from which our "collections" were extracted for modeling. For example, Collection 1 is a subset of data from our humanities_keywords dataset. (See our Datasets page for more information and access to our Zenodo deposits.)

But how did we "read" hundreds of thousands of newspaper articles or millions of social media posts at a time in any of our collections to understand "what everyone says" about the humanities?

We used a method of computational machine learning called “topic modeling.” Topic modeling helps us find major themes of discussion in the media, learn their relative importance, and identify specific articles and sources highly associated with those themes. Widely used in the sciences, social sciences, and recently the humanities (in the new field of “digital humanities”), topic modeling is one of the ways computers can now assist humans in understanding large collections of texts and other materials.

Topic modeling is an unsupervised machine learning method, meaning that it does not presume that a researcher already has a hypothesis or result in mind that “supervises” what the computer is asked to find by prioritizing results that match a predefined example set.Topic modeling works by discovering in large numbers of texts what appear to humans to be thematically coherent “topics.” It does so by analyzing the statistical co-occurrence of words. That is, it finds out which words tend to appear together in a set of documents (and in individual texts).

An article about a political election, for example, might frequently use words naming a nation’s capital city (e.g., "London”) together with such other words as "Parliament." In a topic model, a list of the most frequent words in such a co-occurring group of words can often suggest that the articles sharing many of such words participate together in a "topic"—in this case, one that might be labeled something like "British government."

Topic modeling not only identifies topical themes but indicates their relative weights. (For example, a topic about economics might be statistically heavier than one about the humanities in a particular collection.)It also shows which specific documents participate strongly in a topic (or in several topics at once, since an article mentioning "London" might talk partly about politics but also partly about economics because London is also a financial capital).

Topic modeling thus not only helps find gross patterns but guides human readers to specific examples of documents to examine closely because of their statistical association with a topic. Such guidance is an improvement on cherry-picking examples anecdotally or because they fit a researcher’s preconceived "hunch" about what is going on.

Materials: Our collections can be downloaded as “bags of words.” (This means that except for open-access documents or social media, we are constrained by copyright to sharing original documents only as word frequency counts and other "non-consumptive use" data, which can be used for computational modeling methods such as topic modeling but not for ordinary human reading).

Tools: You can download our Workspace of Jupyter notebooks and other tools for performing topic modeling our texts (or your own) and then for visualizing models in the interactive interfaces of our Topic Model Observatory. The Workspace is available in a Docker “containerized” form [soon to be released] so that you can reproduce everything on your own computer and run it. We also make available our Topic Model Interpretation Protocol—standard instructions and observation steps for researchers using topic models.

(See WE1S GitHub and Zenodo deposits for access to tools we have so far released under the open MIT license.)

However, computational machine-learning methods like topic modeling (and such other methods as supervised text classification and "word embedding" we also used in a supplementary way) have been called “distant reading” for a reason.While topic modeling may guide humans to particular articles for “close reading,” there is still an inescapable sense for humans of great distance---as if seeing things from outer space, because that is how far away we have to go to see “what everyone says.”

That's why we called our suite of adapted and original interactive visualization interfaces for exploring data models our "Topic Model Observatory."

So besides closely reading articles that our topic models pointed us to, we added an up-close human touch. We surveyed students and others, and met them in focus groups in which we discussed their views of the humanities.

Our “human subjects research”—as it is called by the “institutional review boards” at our universities, which for ethical reasons must approve research involving people—was small-scale and local by comparison with the Humanities Indicators U.S. survey. We conducted it at two of our three campuses partnering on the WE1S project: U. California, Santa Barbara, and U. Miami.

But the human subjects research was illuminating. The purpose was to give us a gut check on what the media says about the humanities. We needed to hear what students and others near us were saying. Our local and small-scale results helped us better understand the wider media discourse by seeing the humanities from various specific perspectives—for example, the student viewpoint (including that of our project institutions' many first-generation college or first-generation immigrant students).