Data-Efficient Playlist Captioning With Musical and Linguistic Knowledge

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11401 - 11415 December 7-11, 2022 ©2022 Association for Computational Linguistics Data-Efficient Playlist Captioning With Musical and Linguistic Knowledge Giovanni Gabbolini Insight Centre for Data Analytics School of Computer Science & IT University College Cork, Ireland giovanni.gabbolini@insight-centre.org Romain Hennequin and Elena Epure Deezer Research Paris, France research@deezer.com Abstract Music streaming services feature billions of playlists created by users, professional editors or algorithms. In this content overload sce- nario, it is crucial to characterise playlists, so that music can be effectively organised and ac- cessed. Playlist titles and descriptions are pro- posed in natural language either manually by music editors and users or automatically from pre-defined templates. However, the former is time-consuming while the latter is limited by the vocabulary and covered music themes. In this work, we propose PLAYNTELL, a data- efficient multi-modal encoder-decoder model for automatic playlist captioning. Compared to existing music captioning algorithms, PLAYN- TELL leverages also linguistic and musical knowledge to generate correct and thematic captions. We benchmark PLAYNTELL on a new editorial playlists dataset collected from two major music streaming services. PLAYN- TELL yields 2x-3x higher BLEU@4 and CIDEr than state of the art captioning algorithms. 1 Introduction Playlists are a popular feature of music streaming services. Users consume playlists for the 31% of their listening time (Schedl et al., 2018). And, 55% of users create their own playlists (Muligan, 2017). Playlists are also created for users by professional editors or algorithms. For instance, the popular music streaming service, Spotify, was hosting more than four billion playlists in 2021(Dean, 2021). In this content overload scenario, it is crucial to characterise playlists, so that music can be effec- tively organised and accessed (Choi et al., 2016). A common approach is to rely on automatic playlist tagging with tags such as music genres or decades. However, tagging solutions are limited because they often resort to a pre-defined set and fall short of filling the semantic gap between audio and human-like descriptions (Choi et al., 2020). In contrast, natural language could be used to char- acterise playlists with less ambiguity and using a richer vocabulary. In fact, curators often provide a title and/or a description of their created playlist in natural language. However, describing playlists, es- pecially the ones created by recommendation algo- rithms, is very labor-intensive and time-consuming when done manually (Doh et al., 2021). When au- tomatic captions are proposed, these often rely on pre-defined templates thus cannot cover all kinds of cases, similar to tags (Afchar et al., 2022). To address the above limitations, Choi et al. (2016) introduce the task of playlist captioning, that is automatically describing a playlist using natural language. Playlist captioning can enable several useful applications, such as assisting cura- tors in the process of finding an appropriate caption for a playlist; enabling search and discovery of playlists through human-like queries (Manco et al., 2021); assigning captions to algorithm-generated playlists, that could be also used as explanations for automatic …

Content truncated to 500 words