P1-8: Toward Song Exploration Without Listening Through Lyrics-Derived Linguistic Information
Kanako Suzuki, Kosetsu Tsukuda, Takayuki Nakatsuka, Kento Watanabe, Tomoyasu Nakano, Masataka Goto, Keitaro Tanaka, Shigeo Morishima
Subjects: User-centered evaluation ; Cognitive and user-centered MIR ; Applications ; Lyrics and other textual data ; MIR data fundamentals ; Music retrieval systems
Presented Virtual
Music streaming services typically assume that users explore songs while listening to them, which often exposes users to songs that do not match their preferences. Although lyrics are an important cue for judging songs, reading the full lyrics of many songs during exploration imposes substantial cognitive load. In this paper, we investigate whether lyrics-derived linguistic information can support song exploration without listening. Specifically, we use a large language model to generate 17 types of linguistic information from song lyrics, including a catchphrase, a motif, and a summary text, and conduct a crowdsourcing-based user study involving 1,235 participants. Our analysis results provide several reusable insights, including the following: (1) There is substantial demand for exploring songs by examining various kinds of information so as to avoid listening to songs that do not match users' preferences. (2) When lyrics-derived linguistic information is judged as "Dislike," the full lyrics are also often judged as "Dislike." (3) Combining multiple types of linguistic information increases preference consistency with the full lyrics while requiring much less text than reading the full lyrics. We also discuss how these insights can be applied to song exploration systems.
The changes from the initial submission are the following three points.
- In response to a reviewer's request for further clarification regarding VOCALOID songs and Nico Nico Douga, we added the following footnote in Section 4.1.
On Nico Nico Douga, it is quite common for both amateur and professional musicians to upload songs created with VOCALOID.
- We removed the following statement regarding the supplementary material submitted for review from the footnote in Section 3.
The prompts used to generate each of the 17 types are described in the supplementary material.
- To meet the page-length requirements, we moved the description of the LLM from the footnote in Section 3 to the main text as follows.
The linguistic information for each type was generated by using an LLM (the API of GPT-5.1),