In this work, we implemented a small tool based on deep learning system with information retrieval function, which generates rap lyrics by rapper names or keywords given by users.
To be specific, we collected and clean lyrics of existing hip-hop songs as a training data set at first. Next we extracted keywords and do sentimental analysis for each piece of lyrics, and recorded both to create inverted index. Then trained a LSTM model to generate new rap lyrics based on the query from users.
One of the most challenging problem nowadays is due to the difficulties from combining the emerging technologies and people’s increasing requirements over all kinds of entertainment, such as music, which can’t be explained by digits and letters mathematically.
Over all kinds of music, rap is the one which requires closest connection not only between the lyrics content and the melody like others, but also between the lyric's pronunciation and the rhythm. Lyrics, especially rap lyrics, tend to contain colloquial expressions to establish a flow and rhyme scheme throughout a song[1]. So far, many different approaches have been proposed to generate music pieces by artificial intelligence by certain algorithms but few of them explore the rap due to its unique attributes.
There are two main points to address this problem. The first one is that the processing of rhyme scheme which need to be highly embedded into specific melody naturally. Second, the automatically keyword extension must obey both grammar and melody’s emotional environment, known as sentiment, which veto embedding words straightforward into melodies. For instance, the keyword given, ‘wedding’, would be extended by proposed framework to related complete lyrics and the melody chosen for it should also with the similar sentiment.
Furthermore, the evaluation process needs to be put into consideration. Unlike mathematical data which possesses fixed and obvious label and variables for evaluation, this work which is related to natural language processing need to consider different aspects of the generated results. In this project, we aim to follow the rules above and propose a framework that geared towards imitating musician’s creation and regenerate new rap by trained model with few given keywords.