Facebook Papers about bridging
A number of Facebook Papers — documents leaked by whistleblower Frances Haugen in 2021 — describe internal research efforts conducted at Meta which are relevant to the idea of bridging. The papers can be accessed via fbarchive.org, an online repository managed by the Public Interest Tech Lab at Harvard. (Note that you will need to create an account with them first and log in to fbarchive.org before the links below will work.)
Below, we list excerpts from some of the most relevant documents.
Diverse positive motifs
The following important work on the value of diverse positive motifs as ranking signals was carried out by Matt Motyl and others.
How Diverse Engagement May Identify Valuable Civic Comments
Text analysis indicated that comments with more positive reactions from diverse audiences contained more language emphasizing relationships with other people, helping others, & expressing beliefs in a more tentative (vs. dogmatic) fashion. Additionally, these comments contained less harassing language, less ridicule, & less negativity.
Ad hoc labeling indicated that such comments were also perceived as better for the community, higher quality, more informative, and containing fewer attacks.
…adding diversity to comment ranking may enhance our ability to mitigate echo chambers and reduce the ease with which inauthentic or malicious actors can game some of our ranking features…
Actor-level histories of posting comments that diverse others react to positively provide insight into who is authentic, high-integrity, and likely to contribute value to the FB ecosystem and promote better user experiences
People with histories of posting comments that get reactions from homogeneous audiences are much more likely to violate community standards, be reported by other users for being fake, bullying, hate, and violence incitement, and have their accounts disabled.
There is an opportunity to improve detection and enforcement against abusive integrity-threatening accounts by incorporating actor-level - histories of generating positive vs. negative engagement from diverse vs. homogeneous audiences.
There is also an opportunity to integrate actor-level - weights into ranking where content from users with a history of generating positive reactions from diverse audiences gets prioritized over content from actors with a reputation for generating negative reactions, or generating reactions from homogeneous audiences.
Diverse Engagement is an Agile, Efficient Lever to Promote Integrity
Comments with diverse negative engagement were:
- 10.9x more likely to be reported
- 3.4-5.2x more likely to have violations
Users with a history of generating more diverse negative engagement were:
- 14x more likely to have bullying & harassing violations
- 5.2x more likely to have hate speech violations
- 3.8x more likely to have V&I violations
- 2.3x more likely to have misinfo violations
- 4.3x more likely to be disabled by civic Abusive Accounts Protocol
In comment ranking experiment boosting based on diverse positive engagement, we saw:
- Increased VPVs (+0.69%)
- Increased VPVs on non hateful comments (+0.3%)
- Decreased VPVs on bullying comments (-2.14%)
Using Actor-Level Histories to Predict Integrity Issues
People with a history of writing comments that get bad reactions from homogeneous audiences are:
- 3.8x more likely to have violence incitement CO violations,
- 5.2x more likely to have hate speech c CO violations,
- 14x more likely to have bullying & harassment CO violations, and
- 2.3x more likely to have misinfo CO violations
relative to people with a history of writing comments that get good reactions from diverse audiences.
Diverse Positive Motifs Can Improve Civic Conversations
High level summary of the above work, and future directions.
Polarization
Toward a Composite Behavioral Polarization Metric
The first challenge we faced in this effort was defining in-groups and out-groups in a way that didn’t rely on hard categories or sensitive data. You can read more about news2vec, an unsupervised method for measuring alignment between individuals based on their news engagement patterns, here.
We recommend tracking negativity and incivility minus positive cross-cutting civic interactions as a composite measure of polarization. Consistent with the CG&P team’s principles, this will encourage teams to find ways to reduce negative expressions of polarization, and promote more constructive ways of interacting with different groups.
FB vs. Polarization project demo
We built a new metric, Crosscutting Interaction Value (CIV), to uncover where ideologically separated users are having positive interactions.
Incentives
Political Party Response to '18 Algorithm Change
Political parties across Europe claim that Facebook’s algorithm change in 2018 (MSI) has changed the nature of politics. For the worse. They argue that the emphasis on “reshareability” systematically rewards provocative, low-quality content. Parties have always maintained a mix of positive and negative content, but they feel that they have been forced to adapt to the change by producing far more negative content than before. Engagement on positive and policy posts has been severely reduced, leaving parties increasingly reliant on inflammatory posts and direct attacks on their competitors. Many parties, including those that have shifted strongly to the negative, worry about the long-term effects on democracy.
Posts with Negatively Charged Comment Threads Fare Better in Feed
There’s a (visible) general correlation between negative comment sentiment and number of outbound clicks (imperfect proxy for VPVs). From a publisher’s point of view, this data would seem to encourage posting more content that leads to negatively charged comment threads.