by John Higgins, February 2015
On October 30, 2014, SFMOMA and Stamen Design hosted Art + Data Day at the new Gray Area Art and Technology Theater in San Francisco. The event was formatted as an “unhackathon,” focusing on collaboration and problem solving — rather than competition and speed — as a way of testing an alpha version of SFMOMA’s new API. Since public sharing is a focus of the API, all code created during the event has since been posted to GitHub, and all of the projects have been summarized on SFMOMA Lab. This final installment from Art + Data Day addresses the Art Sentiment Graph developed by SFMOMA’s lead software architect, John Higgins.
The primary experience with a work of art is emotional, and the purpose of artistic expression is, arguably, to elicit that emotional experience. Working remotely, I investigated whether the titles artists give to their artworks accurately frame the emotional responses that the works intend to evoke.
Sentiment analysis algorithms are used to find and extract subjective information found in text, with the aim of determining the attitude of the writer. This technique, although imperfect, is commonly used by businesses to identify the sentiment of social media comments or reviews, making positive sentiment a virtual currency for marketers. The result is the diminished context and nuance of language, which for these purposes is distilled to a positive or negative numeric value. If artworks, like advertisements, are intended to spark an emotional response, then applying sentiment analysis to the titles of an artist’s artworks will generate an additional data metric. Although this metric itself may be subjective in what it tells us about the artwork or the artist, it does reveal the artist’s tendency over time to express a more positive or more negative sentiment.
The program I wrote generated a graph of the sentiment analysis for all artworks by a chosen artist in SFMOMA’s collection. I queried whether artists tended to title their artworks more positively or negatively over time. Would a series of positively titled artworks be followed by a corresponding number of negative ones? Was there any correlation between where the artist was from geographically and the sentiment of the artwork titles?
Two APIs were used to build the sentiment analysis: the SFMOMA Collection API and an API to a sentiment analysis engine. The Collection API was used to get a list of titles and dates for an artist’s artworks and to pass each title to the sentiment analysis engine API, which in turn responded with numeric values between -1 and +1.
The most negative value for a piece of text is -1. For example, The Lonely Metropolitan, a photomontage piece by Herbert Bayer, scores -0.92, a satisfyingly high negative score. Disturbingly, however, the installation titled Pornography in the Classroom [archive master for monitor, “Prick”] registers a highly positive score of 0.76.
Sentiment analysis is programmatically logical and thus incapable of detecting the more complex sentiments of an artwork title, such as Skull of a Gorilla, a visually dark and complex painting by Francis Bacon. Sentiment analysis gives Skull of a Gorilla an indifferent score of 0, in contrast to the truly arresting emotional response elicited by the visual experience of this artwork. I soon also discovered that many artwork titles, such as Untitled or numbers in a series or edition, were scored as neutral by sentiment analysis, and thus returned a numeric value of zero. According to sentiment analysis these neutral titles have no currency.
The sentiment analysis graph page is built entirely from front-end code, JavaScript, and some enhancements via JQuery to help get data from the two APIs. Developed quickly and with a limited selection of only the first 100 artists from the Collection API, the art sentiment graph provides a framework for engagement and spurs additional curiosity. It is a wonderful example of how a clean API can facilitate sketching with code. Future developments to the sentiment graph might make it much more interactive or possibly display an image of each artwork graphed. As the Collection API can return geographic information, this data might be used to generate a map of sentiment data and reveal whether some areas reflect greater negativity or positivity.