[Paper Review] Youling: an AI-Assisted Lyrics Creation System

Zhang Rong-sheng, Xiaoxi Mao+7 more|arXiv (Cornell University)|Jan 18, 2022

Topic Modeling13 references4 citations

TL;DR

Youling is an AI-assisted lyrics creation system that supports interactive, multi-turn generation and iterative revision of lyrics using multifaceted control over content and format. Built on a fine-tuned GPT-2 model with 210M parameters, it enables music creators to generate, refine, and polish lyrics through user-guided sentence and word-level editing, enhancing creativity and control in a human-centered workflow.

ABSTRACT

Recently, a variety of neural models have been proposed for lyrics generation. However, most previous work completes the generation process in a single pass with little human intervention. We believe that lyrics creation is a creative process with human intelligence centered. AI should play a role as an assistant in the lyrics creation process, where human interactions are crucial for high-quality creation. This paper demonstrates extit{Youling}, an AI-assisted lyrics creation system, designed to collaborate with music creators. In the lyrics generation process, extit{Youling} supports traditional one pass full-text generation mode as well as an interactive generation mode, which allows users to select the satisfactory sentences from generated candidates conditioned on preceding context. The system also provides a revision module which enables users to revise undesired sentences or words of lyrics repeatedly. Besides, extit{Youling} allows users to use multifaceted attributes to control the content and format of generated lyrics. The demo video of the system is available at https://youtu.be/DFeNpHk0pm4.

Motivation & Objective

• To address the lack of comprehensive AI-assisted tools for music creators in lyrics generation.
• To shift from single-pass, model-centric generation to a human-centered, interactive collaboration model.
• To support rich control over lyrics content (e.g., emotion, theme, keywords) and format (e.g., rhymes, acrostics, line count).
• To enable iterative revision at both sentence and word levels for high-quality lyric refinement.
• To demonstrate a practical, user-friendly system that integrates generation and editing in a single workflow.

Proposed method

• The system uses a GPT-2-based language model (210M parameters) fine-tuned on 300K Chinese lyrics for high-quality generation.
• It supports two generation modes: full-text generation (one-pass) and interactive generation (context-conditioned, line-by-line).
• Users input multifaceted attributes, including content controls (text style, emotion, theme, keywords) and format controls (rhyme, acrostic, line count, word count per line).
• The revision module uses a sequence-to-sequence framework trained on masked lyrics, where [MASK] tokens replace target sentences or words for prediction.
• Training examples are created by randomly replacing sentences or words with [MASK], with style tags as context for the revision model.
• Version control is implemented to allow users to safely iterate and revert changes during the editing process.

Experimental results

Research questions

• RQ1 How can AI systems support a human-centered, interactive lyrics creation process rather than relying on single-pass generation?
• RQ2 What types of multifaceted control attributes are most effective for guiding meaningful and stylistically coherent lyrics generation?
• RQ3 Can iterative revision at the sentence and word level significantly improve the quality of AI-generated lyrics?
• RQ4 How does integrating generation and editing in a unified system enhance user creativity and efficiency?
• RQ5 To what extent can a GPT-2-based model, fine-tuned on lyrics, support diverse control and high-quality output in real-world music creation?

Key findings

• Youling supports the largest variety of content and format controlling attributes among existing systems, including text style, emotion, theme, keywords, rhymes, acrostics, and structural constraints.
• The interactive generation mode enables users to select from multiple candidate lines conditioned on preceding context, improving control and quality over one-pass generation.
• The revision module successfully supports both sentence-level and word-level editing, allowing users to iteratively refine lyrics until desired output is achieved.
• The system demonstrates that fine-tuning a 210M-parameter GPT-2 model on a 300K Chinese lyrics corpus enables high-quality, controllable generation with strong relevance to input attributes.
• The integration of full-text generation, interactive generation, and revision in a single interface enables a seamless, user-friendly workflow for music creators.
• The system is publicly available at https://yl.fuxi.netease.com/ with a demo account, enabling real-world testing and feedback.