Learning from Conversation Trees
My name is Bridgit Mendler and I am an actress, recording artist, and a graduate student at the MIT Media Lab. And I want to tell you why social media — which was an increasingly important aspect of my entertainment career — is now a focus of my research.
As an entertainer with a large following, I have had experiences on social media when my posts travel pretty far, pretty fast. Sometimes this is inspiring, sometimes this is paralyzing.
Recently I put out a simple tweet addressed to the internet void: “how are you guys?” One person told me she wasn’t in school because she had been protesting in her country for over fifty days. Another person told me he is writing a book of poetry. Other people were celebrating being done with exams. Others were struggling with anxiety and depression. There were hundreds of messages in minutes. I kept scrolling and scrolling down the feed, seeing the responses from people in dozens of countries.
Conversation Trees
I came to the Media Lab wondering if there was a way to understand this influx of information. To my knowledge, the only existing social media tools were about “driving engagement” and “maximizing sales.” It was all about more without even knowing what more consisted of. I wanted to know whether a tool could help me make sense of the mass of conversations on twitter with coherence and, hopefully, meaning.
That is why, with a number of collaborators (Peter Beshai, Prashanth Vijayaraghavan, Soroush Vosoughi, and Ann Yuan) I have built a prototype for mapping conversations at scale online that I call “Conversation Trees.” This approach allows me to look at the comments in reply to one tweet and see, structurally, how people are talking to each other. These trees are all conversations on Twitter that have roughly the same volume of replies, but the structure of replies is significantly different.
Do you want to know what happens to a Kanye tweet once it hits Twitter?
The node at the top is a tweet from Kanye. At the first level down from that node are other nodes that represent tweets in direct response to Kanye. The tweets below that are in response to those responses and so on and so forth until you have this “tree” structure. The colors reflect the emotional content of the tweets, with different colors representing an angry tweet or a happy tweet or a sad tweet, etc.
This is the way I have been visualizing conversations. My hope is that, with the help of tools like Conversation Trees, we can find robust patterns of conversation and people can better understand their impact online. Through my time at MIT and through familiarizing myself with the research of others around me, I have been inspired me to think more broadly about what the value is in better understanding online discourse.
I am currently in a master’s program at the Media Lab’s Laboratory for Social Machines (LSM), which aims to promote deeper learning and understanding of human networks through data science. Some of the research at LSM funnels into Cortico, a non-profit that’s seeking to build a healthier public sphere and deploy tools based upon our research.
My projects are at the overlap between LSM and Cortico, and are centered on what we call “public sphere health metrics.” We conceive of the public sphere as all forms of discourse, from global social media platforms to local community storytelling. By “health metrics” we simply mean measuring how well the digital public sphere is enabling positive and productive discourse.
One hypothesis we have is that incivility online tends to inhibit productive discourse. People who study social media talk a lot about echo chambers and filter bubbles. I believe the harm in echo chambers and filter bubbles is that we are not well-positioned to interact with people who are different from us. The world is full of many different kinds of people with awesomely diverse opinions, beliefs and needs. And sometimes they clash. Lately, they seem to be clashing frequently and intensely. At LSM, we believe that improving the health of public sphere discourse starts by measuring and tracking it first.
I’m now working with other researchers on measuring tweets to detect patterns in structure and content. We are classifying content by looking at the emotion and civility of comments like those in the Conversation Trees above.
People use language in complex ways; any analysis of online discourse should reflect that complexity. Algorithm-based machines, as I’ve found through my work at LSM, are kind of literal. For instance, if someone says “shut up” in a tweet, our tools would most likely register that tweet as angry and uncivil. If someone says “that’s great,” our tools would most likely register that tweet as positive and civil. But people can use these same phrases to communicate vastly different things. Let me remind you of Regina George, the leading mean girl in the wildly popular movie, Mean Girls. When she says “shut up” in the movie, it is an expression of enthusiasm and excitement. If we were to annotate the language of Regina George correctly for emotion, we would need to label “shut up” as an expression of happiness. But what about sarcasm? When someone says “that’s great” they might really mean the opposite, as in “wow, that sucks.” This is why, when it comes to labeling the emotional content of tweets, context is so important.
Many people use angry language or expletives to communicate something positive online. And many people use “positive” words to communicate something rude or offensive. If you know the surrounding context of the posts, you would better understand the intended tone. But how can a machine understand?
Right now, machines aren’t sophisticated enough, and thus social media engagement metrics are shallow. The basic assumption today is that more engagement = good, no matter what. But what if you get high engagement on a tweet that is actually spreading anger and negativity across the Internet? If I had tools as a social media user that accurately made me aware of the impression I leave when I’m posting online, I might make different choices. Or I might make the same choices for different reasons.
What if you were able to see that you started a thoughtful and productive discussion versus a runaway argument of explosive insults? Would it be worthwhile to know that you hurt the feelings of ten thousand people or, alternatively, that you made their day? Machines cannot interpret these things on their own.
I think understanding impact requires a lot of public collaboration and participation. Through my research, I would like to create a shared process with users across many demographics and audiences to fill some of the gaps remaining in machine-only analysis.
The research evolves and branches, much like the conversation trees themselves. I am excited to welcome more voices into this conversation; to make the understanding of our public sphere more nuanced, reflective, and hopefully more meaningful.
