houdaipha/morad

Implementation of "MORAD: A Multimodal Dataset of Authentic Emotional Expressions in Moroccan Arabic" # MORAD: A Multimodal Dataset of Authentic Emotional Expressions in Moroccan Arabic This repository hosts the code for evaluating models on MORAD, a multimodal dataset that captures authentic emotional expressions in the Moroccan Arabic dialect. ------------ ## MORAD corpus MORAD is a multimodal dataset created to support emotion recognition research in the Moroccan Arabic Dialect (Darija). The dataset includes: - 2,325 videos totaling 190 minutes (3 hours) of content. - 329 unique speakers expressing emotions in unscripted, natural settings. - Seven emotions: neutrality (44%), anger (22%), sadness (15%), happiness (14%), fear (2%), disgust (2%), and surprise (1%). - Three modalities: acoustic, visual, and textual. To enable flexible evaluation, results are reported under two configurations: - **4-way setup**: covering happiness, sadness, anger, and neutrality. - **7-way setup**: extending the 4-way set with fear, disgust, and surprise. Various deep learning models were evaluated on MORAD, including unimodal (speech, visual, and text) and multimodal approaches. The highest performance was achieved by the multimodal model (CLIPER) with a weighted F1 score of **73.11%** and an accuracy of **73.36%** on the 4-way task, and **66.69%** and **69.38%**, respectively, on the 7-way task. *Figure: Overview of the MORAD dataset collection pipeline.* *Figure: Word cloud representation of the filtered most common words in MORAD.* ## Evaluation Results We benchmarked several emotion recognition methodologies on the MORAD dataset, covering different data modalities. Results are reported using weighted F1-score (WF1) and accuracy (Acc.) under both the 4-way and 7-way classification setups. ### Speech Models | Model | 4-way WF1 | 4-way Acc. | 7-way WF1 | 7-way Acc. | |-------------------------------------------|------------------|------------------|------------------|------------------| | MF-LSTM | 52. …