Interested in the things that people say in the chat as a mediator of the perspectives they bring in
Simple version of this: sentiment analysis on the messages
e.g., “Are Democrats or Republicans more hypocritical?”
Turns out that people’s partisan affect going into the chat is a pretty good predictor of the average sentiment
Affective polarization is more negative
Take something like this to the next level
Ideally draw a through line from pre-discussion treatment to post-discussion positions and attitudes
We are able to observe a lot more of the mechanism than a conventional survey in which you give them a treatment + ask them to answer a survey question
Perhaps better to call it a ‘mechanism’ rather than ‘mediator’
What about accounting for the high-dimensional design space?
Causal persuasion effect mediated by utterances → ‘I put person X in a group with 2 others that are much more conservative’
“Treatment” is manifested by other people saying more conservative thing → conservative utterances had an effect on person X, whose opinions became more conservative than counterfactual had they been placed with liberals
Main Meeting: Summary of Next Steps
Preprocessing (text data)
Dean & Chris - have experience dealing with exactly these challenges, package to help with detecting speakers and segmenting utterances
James has also created a new feature that pushes data to a GitHub repository (instead of AWS)
Something parallel for the deliberation videos could be good
Can get automation set up so that they go through an automation pipeline
Choosing the right features
May want to add new features, like polarization / political affiliation
Modeling the ‘through line’ between inputs to mediator to outputs
Understanding this as a process in which the beginning is connected to the end via the conversation
Can we analyze this in a way to see the dynamic nature of the conversation; such that we can model the way the persuasion is occurring in real time?
Model doesn't have to be perfect; James: perhaps can build towards "prediction competition" in the future to refine model
Emily / Will Pre-Meeting
Interested in the things that people say in the chat as a mediator of the perspectives they bring in
We are able to observe a lot more of the mechanism than a conventional survey in which you give them a treatment + ask them to answer a survey question
What about accounting for the high-dimensional design space?
Main Meeting: Summary of Next Steps