Thoughts on Analyzing and Visualizing Text Messages

Nairobi was shut down during the peak of the pandemic in 2020. The city was closed off from the outside world and a curfew was imposed, limiting movement within. Everyone had experienced this lockdown bogged down in their own estate or community. 

During this time, we deployed an application called Mtaa, a mobile application that enables users to send geolocated text messages with an “emotion” attached to them. We deployed the app to keep in touch with our friends and team members and to collect data on people’s experiences of the pandemic and the measures imposed.

Our friends sent more than 700 messages in a period of only a few weeks. The messages ranged from day-to-day interactions, reports of extreme kindness and sacrifice, to small acts of daily life like painting graffiti and making music videos, to talking about crime and lack of essential services, such as access to water and food. The app gave us a sense of purpose during trying times, but most importantly,  it kept us close and connected through the sharing of experiences. The exercise also gave us an opportunity to see whether we can make these narratives into actionable data. 

Analyzing more than 700 messages posted in several languages represented a unique challenge. There are a few standard ways to analyze such data, either geospatially or statistically, each with its own strengths and shortcomings. I’ll touch on a few in this blog post.   

I’ll start with the most obvious and easiest one: “putting points on the map”. Each message is geolocated, so the most straightforward way is to display them on a map. Points on the map are visually appealing but not particularly useful, especially when points start to cluster in the same area. Once the number of data increases, the map itself becomes crowded and hard to read.

Second, our messaging platform allows people to express their ‘feelings’ about a situation in form of emotions. Users can tag each message based on whether they think it has a more positive or negative connotation associated with it. These so-called ‘feelings’ can be displayed as an aggregate in an area and represented as such. We used hexagons to represent an ‘average emotion’ associated with an area (see the image below). In this way, we can identify areas that have more positive or negative opinions associated with them. However, by displaying opinions as a ‘vibe’ we lose the granularity or detail that makes these opinions so interesting in the first place.  

Third, we can tag each message based on the topic it depicts. For example, if a message talks about water, we tag it as ‘water,’ if it talks about crime, we tag it as ‘crime,’ etc. To do this we have to read through all of the messages, create categories, and manually place each message into a relevant category. Naturally, this is time-consuming but it allows us to look at the statistical and spatial distribution of topics. For example, we can see how many messages there are about the lack of food, security, development, access to services, or even people having a good time, and where they are located. This process could be sped up using language processing techniques but they would have to work on several languages at once. Tagging messages in this way does have its merit in that it allows for easier analysis, however, as in the previous point, we lose the detail and importance of each message as a unique story.

Further, we can experiment with turning people’s experiences into more traditional, social-media-driven narratives by merging visuals with messages. This is a very powerful storytelling technique fit for modern-day social media consumption (see image below), but apart from being visually striking, there’s not a lot there. Plus, it adds human interference through selecting and preparing each ‘experience,’ which interferes with the free-flowing communication about a place that people occupy.

Finally, we can display messages on a timeline. This looks cool but is only useful if we’re interested in the temporal distribution of certain events at a certain location. We can see in the image below that May 8 was a particularly dark day in our collective family. 

I’m sure there are more ways of analyzing qualitative data in the form of text messages. However one goes about it, it is important to note that peoples’ narratives and stories are essential in learning about a place. Telling a story through data needs to be a multilayered process, it has to rely on more than just ‘points on the map’ or raw statistics. We learn the most when we listen with curiosity and when we listen to understand.