Ten Social Dimensions of Conversations
Using AI to find the building blocks of our relationships
Social relationships are paramount. They determine where we work, who we marry, and what we do. That’s why social science researchers— and more recently their “computational social science” cousins — spent considerable efforts to draw systematic categorizations of the fundamental sociological dimensions that describe human relationships. We tried to figure out what are the fundamental dimensions of social relationships and how to detect them automatically from online interaction data.
By compiling an extensive review of decades’ worth of findings in sociology and social psychology, we identified ten dimensions that have been widely used to categorize relationships:
- Exchange of Knowledge 📚
- Power dynamics (e.g., between a boss and their employee) 👑
- Status giving — conferring appreciation, gratitude, or admiration 🙏
- Expressions of Trust 🤞
- Giving emotional or practical Support 🤗
- Romance 🌹
- Similarity — shared interests, motivations or outlooks 👯♂️
- Identity — Sense of belonging to a group or community 🏳️🌈
- Fun 😂
- Conflict ⚔️
It turns out that people find these dimensions very relatable. Through a small crowdsourcing experiment, we asked 200 people to spell out words that described their social connections and found that all of them fitted into the 10 dimensions.
Being able to quantify these 10 dimension from interaction data is quite useful because, by combining them in opportune proportions, one can draw an accurate, explainable, and intuitive description of the nature of most relationships.
In our work that will be presented at The Web Conference 2020, we used Natural Language Processing to measure the 10 dimensions from conversational text. We asked crowdworkers to label 9,000 Reddit comments with the 10 dimensions. We used this data to train AI tools that can predict all dimensions rather accurately from potentially any text— for 🤓 eyes: using BERT and LSTM we got AUCs ranging from 0.75 (Identity) to 0.98 (Romance).
And now, the fun part. We applied our tool to online data discover how the 10 dimensions are used in different contexts and at different levels of aggregation: individual messages, relationships, groups, and even entire States in the US!
Social Dimensions in Individual Messages — movie scripts 🎥
Scripted movie dialogs are fictional yet plausible representations of conversations that span a wide spectrum of human emotions and relationship types. We ran our models on all movie lines from the Cornell movie script corpus, and found out what are the dimensions conveyed by iconic movie lines.
Next, a few lines from some iconic 20th century movies. We reported the detected levels of the 10 dimensions in the inset diagrams; those above the dashed line are the most significant ones.
Ben transfers his knowledge of the ways of the Force to Luke. This line highlights the power relationship between the master and the pupil, and warns about impending conflict.
Michael likes to think of himself as a man of honor. He tributes appreciation (Status) to his friend for his services, and grants him Trust in the wake of suspicions of betrayal.
It seems that the supercomputer HAL 9000 is trying to provide some technical support (= knowledge + support) to an increasingly worried astronaut.
The identity of a Vulcan, expressed before the extreme sacrifice 🖖.
Social Dimensions in Relationships — Twitter 🐦
Dimensions estimated at message-level can predict the dimensions that people would use to describe their social relationships. To show that, we collected data from tinghy.org, a platform of games that gathers people’s perceptions about their Twitter friends (see pic on the left). We found that we can get good predictions of relationship-level labels (e.g., a relationship based on trust) by looking at 20 twitter replies or more between the two people involved.
Social Dimensions of Communities — Enron corp. emails 📧
Enron Corporation was an American company founded in 1985 that went bankrupt in 2001, when its systematic practices of accounting fraud were exposed to the public. After the scandal and the resulting investigation, The Enron Email Dataset was released to the public. We looked at how expressions of different dimensions varied over time in Enron emails, on average.
The evolution of conversational dimensions reveals a rich picture that matches the known stages of Enron’s downfall. As the initial concerns sparked, the exchange of status and support plummeted: panic started to spread and employees stopped celebrating their achievements, thanking each other, and offering comfort. About three months later, the frequency of knowledge exchange dropped sharply: as concerns grew, employees spent less time in dealing with their everyday duties. A few weeks before the layoffs, as it became clear that many employees would have been made redundant, conflict exploded and the power structure collapsed — fewer orders were given to the angry crowd of employees who were made aware of the impeding jobs cuts. In the aftermath of the layoffs, those who managed to stay in the company gave support to each other for a few weeks before the imminent crack.
Social Dimensions in Society — US States statistics 🦅
Last, we tested whether the presence of those dimensions in conversations is associated with real-world outcomes at societal level. We extracted the ten social dimensions from 160M Reddit messages posted by 1M users who we could geo-reference at the level of US States. We conducted a geographical analysis to study the relationship between the presence of the 10 dimensions and socio-economic outcomes, estimated by official statistics. In particular we checked the relationship between:
- Exchange of knowledge and average education level
- Exchange of knowledge and average income
- Expression of social support and suicide rates
As expected, Knowledge is positively associated to education levels and to wealth — knowledge exchange is a driver of innovation and economic growth. Support is positively associated with suicide rates: people affected by depression, especially those who have suicidal thoughts, do not tend to trust their peers and seek social support in different contexts, often online.
All the dimensions combined predict the official statistics quite accurately (up to an adjusted R2 of 0.52), even after discounting for confounding factors like population density.
What is this all for?
The ability of automatically extract fundamental dimensions of social exchange from text could contribute to creating research and analytics tools for social networks. We believe that the dynamics of a number of processes mediated by social networks could be re-interpreted with our application of the 10 dimensional model to conversation networks. For example, researchers could look at the phenomena of information diffusion, spreading of fake news, polarization, and link creation in the light of the ten social dimensions.
Social media companies could monitor the prevalence of the ten dimensions in the public online discourse as a way to detect and promote positive and meaningful interactions —for example those based on support and trust — rather than just punishing misbehavior.
Want to learn more?
Check out the preprint or our data and code. Research conducted by @minje_choi at @umsi and @lajello, @varkrisz, and @danielequercia at @BellLabs. Please contact us for any questions!
