LyCon: Lyrics Reconstruction from the Bag-of-Words Using Large Language Models (2408.14750v1)

Abstract: This paper addresses the unique challenge of conducting research in lyric studies, where direct use of lyrics is often restricted due to copyright concerns. Unlike typical data, internet-sourced lyrics are frequently protected under copyright law, necessitating alternative approaches. Our study introduces a novel method for generating copyright-free lyrics from publicly available Bag-of-Words (BoW) datasets, which contain the vocabulary of lyrics but not the lyrics themselves. Utilizing metadata associated with BoW datasets and LLMs, we successfully reconstructed lyrics. We have compiled and made available a dataset of reconstructed lyrics, LyCon, aligned with metadata from renowned sources including the Million Song Dataset, Deezer Mood Detection Dataset, and ALLMusic Genre Dataset, available for public access. We believe that the integration of metadata such as mood annotations or genres enables a variety of academic experiments on lyrics, such as conditional lyric generation.

• The paper proposes a novel method to generate copyright-free full lyrics from BoW datasets by leveraging large language models.
• It integrates rich metadata such as genre, artist, and mood annotations to accurately mirror original lyrical content.
• Comparative analysis shows that the reconstructed lyrics closely match original metrics, enabling advanced lyric studies.

Analysis of LyCon: Lyrics Reconstruction from the Bag-of-Words Using LLMs

The paper "LyCon: Lyrics Reconstruction from the Bag-of-Words Using LLMs" addresses a significant challenge in lyric studies, particularly regarding copyright restrictions associated with using internet-sourced lyrics directly. A novel method was developed to reconstruct copyright-free lyrics from publicly available Bag-of-Words (BoW) datasets, leveraging LLMs and associated metadata, facilitating a way to circumvent these restrictions while preserving the extensive metadata benefits from the original datasets.

Introduction and Motivation

Due to copyright constraints, direct use of internet-sourced lyrics in academic research is limited. Publicly available datasets, such as musiXmatch, offer BoW formats that list vocabulary and word frequencies without providing full lyrical content. However, the absence of complete lyrics limits their utility for research areas requiring full text analysis, such as lyrical structure or generation. The authors introduce a method to reconstruct full lyrics from BoW datasets using LLMs, thereby generating lyrics that align with the original contents in terms of vocabulary, themes, and mood.

Methodology

The methodology integrates metadata from multiple datasets to reconstruct lyrics. BoW data from the musiXmatch dataset, included in the Million Song Dataset (MSD), serves as the core vocabulary source. Additional metadata, such as artist names, song titles, genre information from the ALLMusic Genre Dataset, and mood annotations from the Deezer Mood Detection Dataset, are employed to enhance the quality and contextual relevance of the generated lyrics.

The reconstruction process utilizes OpenAI's GPT-4 model, tasked with generating lyrics based on prompts incorporating genre, artist, title, mood, and vocabulary. The mood is determined using valence and arousal levels in the 2D valence-arousal space (Figure 1). An example of a prompt provided to the model is:

Compose [GENRE] lyrics, in a style reminiscent of [ARTIST] which represents a [MOOD] mood under the title of [TITLE] using the following vocabulary [VOCABULARY].

This approach led to the generation of a comprehensive dataset, LyCon, which includes reconstructed lyrics for 7,863 songs. Each reconstructed entry is mapped to the corresponding MSD song ID, enabling seamless integration with existing metadata.

Dataset and Analysis

The LyCon dataset is compared statistically to the original lyrics, highlighting several key metrics (Table 1). Despite the inherent differences between original and generated lyrics, the reconstructed set shows comparable counts in average words, lines, and sections per song. Interestingly, LyCon exhibits a significantly lower unique unigram count than the original lyrics, suggesting a more repetitive vocabulary. This is a critical observation for further refining the reconstruction approach.

Additionally, the analysis of abstract and concrete words shows a minor deviation in LyCon's aesthetic qualities. While the gap in concrete versus abstract words is small, it indicates that the reconstructed lyrics maintain a similar stylistic quality to the original ones.

Implications and Future Work

The implications of this research are multifaceted. Practically, the creation of the LyCon dataset advances the field by providing a substantial repository of full lyrics that bypass copyright issues. This dataset can support various academic experiments and applications, such as mood-conditioned lyric generation, genre-based analysis, and deeper lyrical studies that were previously infeasible with BoW datasets alone.

Theoretically, the results prompt further investigation into enhancing LLM prompts to generate lyrics that more closely mirror the statistical and stylistic nuances of the original content. The observed discrepancies in unique unigrams and abstract versus concrete words highlight the areas requiring refinement. Future developments could involve integrating additional layers of metadata or employing more sophisticated models to capture the complexities of lyrical compositions better.

Overall, "LyCon: Lyrics Reconstruction from the Bag-of-Words Using LLMs" introduces a valuable resource and methodology for the academic paper of lyrics, setting a foundation for subsequent advancements in this domain. The dataset and its potential applications promise to augment the scope of research in lyric analysis, paving the way for innovative experiments and enhancing our understanding of lyrical artforms through data-driven approaches.