[Paper Review] Love Me, Love Me, Say (and Write!) that You Love Me: Enriching the WASABI Song Corpus with Lyrics Annotations

Michael J. Fell, Elena Cabrio+3 more|arXiv (Cornell University)|Dec 5, 2019

Music and Audio Processing1 references6 citations

TL;DR

This paper presents the WASABI Song Corpus, a large-scale, multilingual collection of 1.73 million song lyrics enriched with automated annotations for structure, emotion, explicit content, topics, and salient passages. Using NLP and audio analysis techniques, the authors extract semantic and affective features from lyrics, enabling intelligent music discovery, recommendation, and analysis, with ongoing model improvements and public release of data and tools.

ABSTRACT

We present the WASABI Song Corpus, a large corpus of songs enriched with metadata extracted from music databases on the Web, and resulting from the processing of song lyrics and from audio analysis. More specifically, given that lyrics encode an important part of the semantics of a song, we focus here on the description of the methods we proposed to extract relevant information from the lyrics, such as their structure segmentation, their topics, the explicitness of the lyrics content, the salient passages of a song and the emotions conveyed. The creation of the resource is still ongoing: so far, the corpus contains 1.73M songs with lyrics (1.41M unique lyrics) annotated at different levels with the output of the above mentioned methods. Such corpus labels and the provided methods can be exploited by music search engines and music professionals (e.g. journalists, radio presenters) to better handle large collections of lyrics, allowing an intelligent browsing, categorization and segmentation recommendation of songs.

Motivation & Objective

• To create a large-scale, multilingual song lyrics corpus enriched with semantic and affective metadata to support music information retrieval and recommendation.
• To address the challenge of extracting meaningful, structured information from song lyrics, which encode key semantic and emotional content.
• To develop and apply NLP methods for automatic annotation of lyrics with structure, emotion, explicitness, topics, and salient passages.
• To enable music professionals and search engines to perform intelligent browsing, categorization, and personalized recommendations using richly annotated lyrics data.
• To provide open access to the corpus, trained models, and tools for reproducibility and further research.

Proposed method

• Leveraged Web-based music databases (e.g., LyricWiki, DBpedia, Last.fm) to collect and merge metadata for 2.10M songs, with 1.73M containing lyrics.
• Applied a hierarchical segmentation model to decompose lyrics into structural units (e.g., verses, choruses) using sequence modeling and alignment techniques.
• Used a hybrid approach combining structural cues and embedding-based clustering to generate concise, representative song summaries from salient lines.
• Trained a binary classifier to detect explicit content in lyrics using BERT-based embeddings and fine-tuned on a curated dataset.
• Employed a two-dimensional emotion model based on valence and arousal, trained via weakly supervised clustering on lyric embeddings.
• Applied Latent Dirichlet Allocation (LDA) with 60 topics to extract interpretable thematic labels from lyrics, using probabilistic topic modeling on the full corpus.

Experimental results

Research questions

• RQ1 How can song lyrics be automatically segmented into structural components such as verses, choruses, and bridges using NLP techniques?
• RQ2 To what extent can emotion in song lyrics be modeled using continuous valence-arousal dimensions, and how well does this align with human perception?
• RQ3 Can topic modeling effectively extract interpretable themes from large-scale song lyrics, and how do these themes evolve over time?
• RQ4 How do annotations like explicit content and salient passages correlate with genre, emotion, and lyrical structure?
• RQ5 What are the diachronic trends in lyrical themes, explicitness, and emotional expression across decades of popular music?

Key findings

• The WASABI Song Corpus contains 1.73 million unique song lyrics in 36 languages, with 1.05 million annotated with 60-topic LDA models.
• Emotion annotations were successfully predicted for 1.73 million songs using a valence-arousal framework, with 16,000 songs receiving detailed emotion vectors.
• Explicit content was annotated for 715,000 songs, with 52,000 labeled as explicit, indicating a significant proportion of mature content.
• Salient passage summaries were generated for 50,000 songs using a combination of structural and embedding-based selection, improving summarization quality.
• Topic modeling revealed that themes like 'love' and 'drugs' showed distinct temporal trends, with increased representation in certain decades.
• Correlation analysis showed that emotionally intense lyrics were more likely to be structurally segmented and thematically focused, supporting the utility of multi-layered annotation.