We aim to create a “bridging” algorithm that enhances mutual understanding and trust (Ovadya & Thorburn, 2022) in social media interactions. It prioritizes posts that evoke civil discussions while appealing to an ideologically diverse audience. To this end, we rerank content based on 2 criteria: elicited response and diverse engagement. Additionally, we remove toxic posts.
We developed two different models for elicited response: one assessing the intensity of Affective Response (AR) to a post, such as happiness or sadness, and another assessing whether a post attracts Harmful Response (HaR), such as toxicity or offensiveness. These models are trained using the emotional content and toxicity levels of responses, predicted using state-of-the-art pre-trained language models, as variables. Note that a post can simultaneously evoke positive emotions and toxic reactions; therefore, the AR and HaR are treated as separate dimensionsin our analysis.
We measure the diverse engagement of a post through a metric we call "audience diversity” (AD). It estimates the range of ideological slants of the audience engaging with the content (Bhadani et al., 2022). If a post includes a URL, we assess its "source level" diversity by examining the ideological range of the domain's typical audience, which will be pre-calculated using a similar approach as described in Bhadani et al., 2022. Additionally, regardless of whether a post contains a URL, we determine the “topic level” diversity using the textual content.
Our algorithm first removes highly toxic posts, then reranks the rest as follows: Non-HaR posts with high AD scores are prioritized, while all HaR posts are demoted. Where there are ties, posts are ordered by their AR scores to enhance engagement.