Context, images and the limits of digital methods
I am currently at the wonderful Digital Methods Initiative (DMI) Summer School in Amsterdam, working on a Making Climate Social project on the visual language of climate change. More on that soon, but here I want to flag up a key methodological challenge that has emerged for digital methods during the project: that the growing importance of visual communication on social media means research based on keyword search alone is increasingly risky. Relying on keyword searches to research social media platforms may overlook high-engagement posts. I can anticipate the emergence of digital methods that can search images at scale for relevant content. However, a twin-track approach of [digital methods + digital ethnography] is likely to remain optimal for the foreseeable future. Below, I will explain more with reference to a couple of recent examples on Twitter (thanks to Sabine Niederer for work on TCAT).
The US’s decision to withdraw from the Paris Agreement prompted much social media activity, not least on Twitter which is the platform most associated with ‘breaking news’ events. Using the DMI-TCAT tool, we searched for tweets between May 24-June 7, 2017 (approximately one week either side of the US announcement) that included four key terms (climate, climatechange, global warming, globalwarming). The results were then compared to tweets collected manually during ethnographic observation of ‘climate change Twitter’ around the time of the US decision: a wide range of Twitter users with an interest in climate change, covering scientific, policy, political and sceptic communities.
What was immediately clear was some differences between tweets collected digitally (TCAT) and manually (ethnographically). Crucially, some very highly shared content relevant to the Paris Agreement withdrawal did not appear in the digital data, as it did not include any of the search terms as text. For example, French President Macron published a tweet (above) containing no text, just a campaigning meme-like image which contained the slogan “Make Our Planet Great Again”.
Clearly, this was a response to the US government’s decision, and a play on President Trump’s campaign slogan “Make America Great Again”. Interpreting the tweet as related to climate change is dependent not on a particular search term, but on knowledge of a number of contextual factors: the timing (around the Paris decision), the role of Trump (climate change sceptic), the role of Macron (pro-climate action) and Trump’s 2016 campaign slogan. This is important as the Macron tweet was far more shared than any tweet within the TCAT database, with 240,323 Retweets at the time of writing (30/6/17).
One potential digital method for addressing the shortcomings of keyword search is to collect additional tweets from key climate change Twitter users. A researcher with contextual knowledge of the field could set up a list of relevant Twitter users to monitor. In the case above, one might have anticipated that Macron would comment on any US action regarding the Paris Agreement, based on his own presidential campaign. However, another highly shared tweet came from a user one could not reasonably have anticipated. A tweet containing a humorous juxtaposition of memetic text (“Cracking open a cold one with the boys”) and images (Trump with US officials and an iceberg) garnered 62,445 retweets at the time of writing:
Again, this was far more highly shared than anything appearing in the TCAT ‘climate change’ keyword search, but was posted by a user with less than 1,000 followers and no history of tweeting about climate change.
Although this escaped detection using existing digital methods tools, it may be possible to devise methods in the future which could capture such tweets. As with the Macron tweet, there is no obviously relevant text to climate change within the tweet. However, the image content is more directly relevant and might be conducive to a method employing a computer vision API (such as Google Vision), that could detect the image content as Trump and iceberg. The effectiveness of such a digital strategy would require both an improvement in the performance of vision APIs and the expansion of image search terms to a range of possibly relevant terms, such as ‘iceberg’. Even if these challenges were able to be overcome, it is not clear that they would have captured the Macron tweet.
That these two highly-shared tweets were not captured by traditional keyword search methods highlights some limits of digital methods. Of course, digital methods still have an essential role to play in analysing content at scale. However, the examples above demonstrate the importance of visual content and contextual knowledge in social media communication. That the Macron and ‘cold one’ tweets were considerably more popular than anything containing the relevant keywords raises an intriguing possibility: that tweets requiring some contextual knowledge for interpretation have a greater potential for engagement than those containing issue keywords or hashtags. If we consider context and images as part of a broader ‘non-textual’ category of communication (along with videos), then it is clear that we should consider expanding our range of digital methods if social media research is to remain robust and relevant.
