# Detecting Visual Triggers in Cannabis Imagery: A CLIP-Based Multi-Labeling Framework with Local-Global Aggregation

**Authors:** Linqi Lu, Xianshi Yu, Akhil Perumal Reddy  
**arXiv:** [2412.08648](https://arxiv.org/abs/2412.08648)  
**Date:** 2024-11-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.08648v1/

## Key points

- Study title: "Detecting Visual Triggers in Cannabis Imagery: A CLIP-Based Multi-Labeling Framework with Local-Global Aggregation"
- Authors: Linqi Lu, Xianshi Yu, Akhil Perumal Reddy
- Utilized CLIP model to analyze 42,743 images from Facebook (March-August 2021)
- Focus on identifying food-related visuals and analyzing attributes like colorfulness and brightness
- Used BART model to classify ten distinct topics derived from structural topic modeling
- Positive correlations found between certain food-related visuals (fruit, candy, bakery items) and higher user engagement scores
- Positive associations between user engagement and text topics related to cannabis legalization
- Negative associations noted between image colorfulness and specific textual themes within discussions
- Insights for policymakers on implementing warning labels and marketing regulations for recreational cannabis edibles

## Summary

In their study titled "Detecting Visual Triggers in Cannabis Imagery: A CLIP-Based Multi-Labeling Framework with Local-Global Aggregation," authors Linqi Lu, Xianshi Yu, and Akhil Perumal Reddy explore the relationship between visual and textual elements in online discussions surrounding cannabis edibles. The researchers utilized the CLIP model to analyze a dataset of 42,743 images from Facebook during March-August 2021. Their analysis focused on identifying food-related visuals and how attributes like colorfulness and brightness influenced user interaction. They also used the BART model to classify ten distinct topics derived from structural topic modeling and investigate their connection to user engagement levels. Through linear regression analysis, they found positive correlations between certain food-related visuals (fruit, candy, bakery items) and higher user engagement scores. They also observed positive associations between user engagement and text topics related to cannabis legalization. Conversely, negative associations were noted between image colorfulness and specific textual themes within the discussions. These findings provide valuable insights for policymakers seeking to implement effective warning labels and marketing regulations for recreational cannabis edibles. The research highlights how visual cues combined with textual content can impact user engagement dynamics in online discourse about cannabis products.