Music Recommendation for Podcast Scripts: Detecting Emotion from Text

In this project, our task is to generate music recommendations based on podcast scripts. Given a podcast script, our model will detect the emotion from the text to recommend relevant music. To achieve this, our model has been trained on a Twitter dataset with labelled emotions. Then, the model will be trained on podcast scripts through domain adaptation. Initial approaches include building a CNN, LSTM, RNN, BERT, and ALBERT model for emotion detection on the Twitter dataset. In the future, we will perform domain adaptation for podcast scripts and fine-tune our BERT and ALBERT models. Finally, we will have labelled music with emotions, and recommend a random subset of the music with the same predicted emotion of the podcast. 1 Key Information to include • External collaborators: Mercan Topkara (Luminary Media) • Mentor: Dilara Soylu • Sharing project: N/A