How an Alibaba Engineer Built an AI Hip‑Hop Lyric Generator for Double 11

MusicGo is an AI lyric‑generation program developed by Alibaba engineer Lu Yang, designed to compose rap songs for the Double 11 shopping festival by learning from hip‑hop lyrics and technology‑related texts.

Lu Yang describes the core idea as a function similar to Y=WX+B: given an input sequence X, the model should output a sequence Y and hidden state H, then use Y and H as the next input to continue generation.

Deep learning provides the abstraction: a supervised seq2seq model learns the mapping from input sequences (A, C) to output sequences (B, D) by adjusting parameters until the loss is minimized.

Implementation Steps

Collect lyrics – Python scripts scraped thousands of hip‑hop songs from online music platforms.

Preprocess data – removed punctuation and special characters to obtain clean text.

Train a seq2seq model – built an LSTM encoder‑decoder, applied a softmax layer to produce logits, and minimized total loss through iterative training.

Generate lyrics – used the trained model to produce rap verses.

Initial outputs were chaotic because the model had insufficient training and high loss, resulting in incoherent rap lines.

To align the lyrics with Double 11 and Alibaba’s smart technologies, Lu Yang fed the model with articles containing keywords such as "Alibaba Cloud", "Cainiao robotics", "Intelligent Customer Service", and constrained the output with hip‑hop rhythm patterns, then manually refined the verses.

Sample excerpt from the final song demonstrates the blend of e‑commerce hype and rap style:

又来到一年 双11 让我买买买 花光所有力气 让我把指挥权交给智能机器

Images illustrate the system architecture and training process: