End User Authoring of Personalized Content Classifiers: Comparing Example Labeling, Rule Writing, and LLM Prompting

Similarly, some users want to categorize their emails into more fine-grained folders other than the default “Spam” or “Promotions” folders [ 63 , 98 ]. As a result, there have been growing calls to empower end users to customize content curation classi-fiers [ 42 , 43 ]. In particular, researchers have argued for personal content moderation tools t... See more

We found that writing prompts generally allowed participants to create personal content filters with higher performance more quickly, primarily driven by having the highest recall despite comparable precision across conditions.

In addition, curation tools should enable easy iteration to improve initial creations incrementally, as so-cial media users may be more amenable to short tasks spread out over many sessions as opposed to s single lengthy task. Instead, existing systems often assume that users want to create highly performant custom classifiers in a single sitting [... See more

As a result, for many users to realistically participate in curation, tools for authoring classifiers should support rapid initialization . As end users often-times engage with internet content as a leisure activity, a successful tool should allow them to quickly and intuitively build an initial classifier with decent performance.

However, many tools for authoring personal classifiers on so-cial media are designed for users with heightened motivation , such as community moderators [ 12 , 40 ] and high-profile content cre-ators [ 41 ]. In focusing on these users, such tools neglect the usage patterns of general social media users, who spend significant time on social media ov... See more