Culturomics: Not Quite Yet

Well, it’s time to stick my oar in on the Google Ngrams discussion. While a number of computational linguistics scholars have pointed out the pitfalls of Google’s latest toy, I think I have a unique perspective to offer on the issue. I understand what the Ngrams creators were trying to do, because I’m trying to exactly the same thing: get some things cooking. My research on contemporary literary reception is not exhaustive or dependent on highly complex statistical models. That’s because literary reception is a huge, multiply mediated field ranging from café conversations to book reviews, and my access to data is limited. But where I have adopted a “core sample” model, choosing a few accessible data sources to make some robust but limited generalizations about readers and reading culture, Google has gone for the moon shot. By creating an opaque front-end to their 5 million book archive, they offer the illusion of a truly global Ngram search—and they emphasize the scale of their ambition by claiming their tool isn’t merely a corpus search mechanism but the portal to a new science of “culturomics.”

As my colleague Matthew Jockers noted in his own oar-insertion post, “To call these charts representations of ‘culture’ is, I think, a dangerous move.” He goes on to suggest it “may be,” but I have to go a bit farther and say “definitely not.” Here’s the problem: we can’t get reasonable, arguable claims about things like culture or literary history unless the limitations of the corpus are acknowledged and dealt with from the outset. Typically, projects like this limit themselves either by going too small or too big, and Google has gone way big. Let me explain what I mean.

Too Small:

The opposite example would be a research project on a small, meticulously tended patch of texts. Classic humanities research, really, but of limited usefulness for making grounded claims about larger literary-historical or cultural issues (at least until enough such small projects emerge with commensurable results that we can begin to construct some causal chains). Traditional humanities as a whole is full of projects that are “too small” for making broad cultural claims because they are limited to a small data footprint. The walled garden of closely tended results is fascinating and lovely to explore, but it’s difficult or impossible to compare the work to anything outside.

Too big:

Google, by contrast, flies off the macro end of the scale by trying to do too much and claim too much. The corpus is amazing, but nevertheless limited and contingent in many ways. As others have pointed out, the OCR is problematic; the metadata is sloppy; the text distribution almost certainly has a number of biases (how could it not? What is the gender, historical and language distribution of the world’s universal library supposed to be anyway?). By choosing to obscure these limitations instead of illuminating them, Google turns “culturomics” into a toy, not a tool.

Fortunately, the data is all there, and these problems can be fixed. Google loves a good algorithm and will presumably figure out solutions to the various technical problems. With luck (and the persistence of its academic research partners) the Ngrams team will also come to acknowledge and reveal the limitations on its data. Once that happens, we can really get cooking and make a clear case for when this vast corpus really does reveal broad cultural trends.

For now, Ngrams is a blunt object but it still has some value as a tool. I’ll post some examples next time.