Algorithmic Gender Bias in Discovery Systems
This is a working draft of a possible future blog post
At GVSU, we used the Summon discovery service to provide search across most of our resources. One of Summon 2.0's features is the "Topic Explorer," a sidebar that provides general reference information when a search is done. The Topic Explorer shows short excerpts from Reference Sources, including Wikipedia, when a search meets certain criteria. From the Summon Press Release announcing Summon 2.0 in 2013:
Developed by analyzing global Summon usage data and leveraging commercial and open access reference content, as well as librarian expertise, this new feature helps users get started (presearch) with the research process and allows librarians to help users when and where they need it most.
The Topic Explorer returns different reference articles about the superhero "batman," for instance, depending on the host library's subscriptions and settings for which reference sources are used. At GVSU, we get a blurb from the Gale Virtual Reference Library:
While The University of Huddersfield has an article from Wikipedia:
Yesterday (November 11, 2015) I tweeted about a search in the Summon discovery service for "stress in the workplace" that brought up the topic "Women in the workforce" from Wikipedia.1 This seemed not only an inappropriate subject to match the search terms, but also one that showed elements of gender bias. I've been following the work of Safiya Nobel, a faculty member in UCLA's Information Studies department. Dr. Nobel's work on gender and racial bias in commercial search engines has been really influential on my own thinking of the ethical responsibilities of libraries that provide digital research tools. I had previously thought that the indexing process of commercial discovery services were not as susceptible as the commercial algorithms that led to the systemic problems in Google's portrayal of women and girls of color. However, seeing this search made me rethink that position, since Summon relies on some algorithmic mechanism to match up these "topic" summaries with searches.
As of this morning (11/12/2015), ProQuest has made a change to either the algorithm or the mapping of topics to remove this particular topic from coming up. While I think that is a good thing, it makes me wonder what other topics might be exhibiting unintended gender or racial biases. In the meantime, I did have the foresight yesterday to run the search in many of the Summon instances listed in the Community Wiki (a small selection of all Summon customers, and sadly, I didn't capture all the listings). I took screen captures of them, and am reproducing them here to show some interesting tendencies (like toggling and English search in non-English Summon instances brings up the "Women in the workforce" Wikipedia Topic). Below are the screenshots:
I'm currently exploring the Topic Explorer's results on a number of potentially problematic search terms pulled from our Summon usage logs. If you'd like to donate anonymous results from your own Summon logs, I'd appreciate the opportunity to broaden the dataset. Feel free to drop me a line at firstname.lastname@example.org or on Twitter at @mreidsma.
Some of the questions I am working on include, but are not limited to:
- Are there other instances of unconscious bias reflected in the algorithms results?
- If so, are they distributed across different reference sources, or are they limited to a single source?
- Are the biases inherent in the subject-to-subject matching alone, or are there instances of bias evident in the content of the Topic panes?
I do want to emphasize that I'm not necessarily putting ProQuest or Wikipedia on trial here by examining these terms (although I am sure some will see it that way; another problem with a library world that has decided to outsource our search and discovery services to commercial entities that do not necessarily share our values.) I'm not looking at the content of the services as much as I am looking at the algorithms themselves.
I don't think that algorithms by necessity have to expose bias, but I also don't think we'll be able to create these kinds of sophisticated algorithms without first examining our own conscious and unconscious biases and making sure that these are not reflected in the way we code.
- I should note that these search terms are used by a colleague of mine to test new search tools, and we had noticed this discrepancy before, at least a year ago. I assumed at the time that this was a quirk in the service that would be worked out as the service matured. After reading Dr. Nobel's work, I was reminded of the search and ran it again and found that the Topic Pane continued to show the biased result. ↩