2025-ai-dev-europe

All resources (slides, code, etc) for AI_Dev Europe 2025: Sarcastically Speaking: Unlocking Multi-modal Sentiment Analysis with NLP and Facial Expressions

Purpose

Please see session presentation and recording for details. The short answer, this is a model for tackling a complex problem and attempting to achieve the best result as possible. The real value is the session recording as this is a demonstration from start (how to tackle a problem) to finish (building the models and doing the multimodal inference). I wrote this session because I dont't think there are any good tutorials/workshops that discuss things like data grooming (at all), how to attack a problem, and decision to make when building a model.

IMPORTANT: It is highly recommended not to use this in production. This is more of a research paper as this and the original research paper has the following limitations:

• the dataset size isn't nearly sufficient for a real-world application (the original research paper is academic only)
• sarcasm is truly a complex problem that will never be perfect despite getting better results (even though 80% is good, we talk about why in the session)

Creating The Facial Landmark Sarcasm Classifier

1. Navigate to demo/mp4_to_facial and follow the README.md to generate the Facial Landmark data
2. Navigate to demo/facial and follow the README.md to create the model and also run inference

Creating The Speech/Acoustic Sarcasm Classifier

1. Navigate to demo/mp4_to_speech and follow the README.md to generate the Facial Landmark data
2. Navigate to demo/speech and follow the README.md to create the model and also run inference

Running the Multimodal Inference

1. Navigate to demo/multimodal and follow the README.md to run inference