Dutch Emotion Detection

NLP & Text Mining module at BUas, Year 2 Block C. Individual research and engineering project focused on low-resource Dutch language processing and transformer fine-tuning.

NLP pipeline for Dutch language emotion classification. Transcribes video/audio via Whisper, classifies 7 emotions using fine-tuned BERTje/RoBERTa. Achieved 84.7% accuracy with SMOTE-balanced datasets.

Key Highlights

01.84.7% classification accuracy across 7 distinct emotion categories in Dutch-language audio.
02.Whisper ASR integration for automatic speech-to-text transcription of video and audio input.
03.Fine-tuned both BERTje and RoBERTa transformer models; RoBERTa outperformed on informal speech.
04.SMOTE oversampling applied to address severe class imbalance in the training corpus.
05.Full pipeline from raw media file to labelled emotion output in a single Python call.