Generating Tips from Song Reviews: A New Dataset and Framework

Jingya Zang, Cuiyun Gao, Yupan Chen, Ruifeng Xu, Lanjun Zhou, Xuan Wang

Abstract

Online music platforms are saturated with user reviews, many of which are long and non-informative, making it difficult for listeners to quickly assess a song's appeal. This paper introduces MTips, the first dataset specifically for music tip generation, and GENTMS, a multi-stage framework designed to extract concise, empathetic, and insightful "tips" from raw reviews. GENTMS combines punctuation correction, stylistic ranking, and diversity-weighted re-ranking. Experimental results demonstrate that GENTMS achieves a top-10 precision of 85.56%, significantly outperforming standard NLP baselines and maintaining robust performance on previously unseen songs.

Introduction

In the digital music era, user-generated reviews are a vital resource for discovery. However, the sheer volume and variable quality of these reviews often hinder rather than help. Most reviews are lengthy, poorly structured, or lack meaningful insight. To address this, we propose the task of "tip generation"—creating short, self-contained descriptions that capture the emotional and artistic essence of a song. Unlike product reviews, music tips are highly subjective and stylistic, often employing metaphors or poetic language.

A music recommendation interface displaying concise, empathetic tips generated from user reviews for songs like "Bad Blood" and "Money".

The GENTMS Framework

The proposed GENTMS (Generating Tips from Music Songs) framework operates in three distinct stages to transform messy review data into high-quality tips:

  1. Punctuation Correction: User reviews often lack proper punctuation, frequently using commas as universal separators. GENTMS employs a BERT-based model to restore correct punctuation, enabling accurate sentence segmentation.
  2. Sentence Relevance Ranking: This stage uses a dual-scoring mechanism:
  3. Content-based Ranking: A modified TextRank algorithm that incorporates "approval counts" (likes), prioritizing sentences that have already resonated with the community.
  4. Stylistic-based Ranking: A BERT-based classifier identifies the "vibe" of a tip, looking for emotional resonance and philosophical depth rather than just factual statements.
  5. Diversity-Weighted Re-ranking: To prevent redundancy, the framework uses a Determinantal Point Process (DPP) and Probabilistic Latent Semantic Analysis (PLSA). This ensures the final selection of tips covers various aspects of the song, such as lyrics, vocals, and atmosphere.

The MTIPS Dataset

To facilitate this research, we developed MTIPS, the first annotated Chinese dataset for music tip generation. It contains 8,003 labeled sentences (3,062 tips and 4,941 non-tips) derived from approximately 50,000 reviews across 128 songs. The dataset spans five diverse genres, including Rock, Chinese Pop, and Absolute Music. Expert annotators labeled sentences based on eight qualitative characteristics, such as sensory artistic conception and background knowledge.

Experimental Results

GENTMS demonstrates superior performance across multiple metrics. On "seen" songs, it achieved a top-10 precision of 85.56%, a 3.34% improvement over vanilla BERT models. Crucially, the system generalizes well; on "unseen" songs, it maintained a precision of 78.89%. Ablation studies revealed that the stylistic-based ranking is the most critical component, with its removal leading to a performance drop of nearly 29%, confirming that the "flavor" of a music tip is essential to its utility.

Conclusion

This work establishes a new benchmark for music-related NLP tasks. By focusing on the unique stylistic and emotional requirements of music reviews, GENTMS provides a practical solution for enhancing user experience on music platforms. Future directions include personalizing tip generation based on individual user preferences and exploring abstractive models that can synthesize entirely new poetic descriptions.

Summary updated

Abstract

Reviews of songs play an important role in online music service platforms. Prior research shows that users can make quicker and more informed decisions when presented with meaningful song reviews. However, reviews of music songs are generally long in length and most of them are non-informative for users. It is difficult for users to efficiently grasp meaningful messages for making decisions. To solve this problem, one practical strategy is to provide tips, i.e., short, concise, empathetic, and self-contained descriptions about songs. Tips are produced from song reviews and should express non-trivial insights about the songs. To the best of our knowledge, no prior studies have explored the tip generation task in music domain. In this paper, we create a dataset named MTips for the task and propose a framework named GENTMS for automatically generating tips from song reviews. The dataset involves 8,003 Chinese tips/non-tips from 128 songs which are distributed in five different song genres. Experimental results show that GENTMS achieves top-10 precision at 85.56%, outperforming the baseline models by at least 3.34%. Besides, to simulate the practical usage of our proposed framework, we also experiment with previously-unseen songs, during which GENTMS also achieves the best performance with top-10 precision at 78.89% on average. The results demonstrate the effectiveness of the proposed framework in tip generation of the music domain.
BibTeX
@inproceedings{zang2023generating,
  author = {Jingya Zang and Cuiyun Gao and Yupan Chen and Ruifeng Xu and Lanjun Zhou and Xuan Wang},
  title = {{Generating Tips from Song Reviews: A New Dataset and Framework}},
  booktitle = {Neural Networks},
  year = {2023},
  url = {https://openreview.net/forum?id=6pG0g9WhoC},
}
Figure