We would like to invite you to the 68th Vienna Deep Learning Meetup: CommentSense.
Agenda
18:30
- Introduction & Welcome by the meetup organizers & hosts
18:45
- Talk: CommentSense: An On-Device AI Browser Extension for Real-Time YouTube Comment Understanding
by Marc Kroll
19:10
- Hot Topics (tbc)
- Announcements
19:30
- Networking & Discussion: Extra Long Edition
~ 22:00 Wrap up & End
Abstract
Online platforms face persistent challenges in moderating user-generated content, as automated systems often miss harmful material while exposing human moderators to significant psychological strain. Sentiment analysis provides a way to enhance user agency by allowing individuals to engage selectively with positive, neutral, or negative content.
This work compares traditional BERT-based models with small language models (SLMs) that incorporate contextual reasoning. Through user studies and model evaluations, we found that rigid, predefined filtering approaches were poorly received, whereas adaptable systems with on-demand labeling were preferred. Three sentiment-classification tools were developed and tested, demonstrating that SLM-based approaches consistently outperformed BERT baselines. Their ability to explain classification decisions improved perceived transparency and trust, while context-aware models achieved higher macro-F1 scores and stronger alignment with human judgment.
These findings suggest that locally deployed SLMs can support healthier digital interactions by reducing exposure to harmful content while preserving user control.
About the Speaker
Marc Kroll is working as AI Product Manager for an Enterprise Content Management Vendor. His main focus is on Search and IDP.
Outside of work Marc is vibe-coding and catching up on AI News.