| Popis: |
Recent developments in music streaming applications and websites have made the music emotion recognition task continually active and exciting. Recognizing the music mood has many advantages. One of them is that it helps find out the prominent feeling or state of mind it brings to a listener. Identifying a listener's taste is the primary motive of currently available music emotion recognition systems such as music streaming systems (YouTube), music recommender systems (Spotify), automatic playlist generation, etc. Being a subdomain of music information retrieval (MIR), for the past years, several challenges of music emotion recognition have been studied and solved by researchers. Music emotion recognition's significant challenges include data accessibility, data volume, recognizing emotionally relevant features, etc. Several researchers have proved that emotionally relevant features can be identified by analyzing multiple features and the semantic features of lyrics and effects of audio signals create music emotion. Then one can initiate the music recognition task from a lyrical perspective because it has semantic features and audio features such as valence and arousal. The challenging part is the availability of these features that influence the music's emotion. The lyrical features relevant for identifying four emotions (happy, sad, relaxed, and angry) were learned with the help of state-of-the-art algorithms. After that, those features were used to predict the feelings of Music4All dataset lyrics from the Music Emotion Recognition (MER) dataset. This work tries to identify the importance of these features through transfer learning and combining pre-trained model BERT. After transfer learning, the BERT model is then applied to the dataset to improve the model's accuracy. The overall accuracy achieved is 92%. |
