Learning to Generate Tips from Song Reviews

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

TL;DR

This paper introduces the first task and dataset (MTIPS) specifically designed to generate concise, empathetic "tips" from music reviews to help users make faster listening decisions. The authors propose GENTMS, a framework that outperforms standard NLP baselines by combining content relevance, stylistic pattern recognition, and diversity-weighted re-ranking.

What Problem Does This Paper Address?

Online music platforms are flooded with millions of reviews. While these reviews are valuable, they are often long, rambling, or poorly punctuated, making it difficult for a user to quickly decide if a song is worth their time.

While "tip generation" exists for products (e.g., Amazon) and travel (e.g., TripAdvisor), the authors argue that music is a unique beast. Music tips aren't just factual; they are highly subjective, diverse, and often rely on figures of speech like metaphors or rhymes. A single song might have "a thousand Hamlets for a thousand audiences." Existing summarization tools fail to capture this emotional resonance and stylistic flair. Furthermore, user-generated reviews often suffer from "comma-splicing" (using commas instead of periods), which breaks standard sentence extraction tools.

Methods

The authors developed a multi-stage framework called GENTMS (Generating Tips from Music Songs):

  1. Punctuation Correction: Before processing, they use a BERT-based model to predict proper punctuation. This allows the system to accurately segment reviews into meaningful "short sentences" rather than fragments.
  2. Sentence Relevance Ranking: This module uses two parallel scores:
  3. Content-based Ranking: A modified version of TextRank that incorporates "approval numbers" (likes). Sentences from highly-upvoted reviews are given more weight, assuming they resonate more with the community.
  4. Stylistic-based Ranking: A BERT-based classifier trained to identify the "vibe" of a tip. It looks for patterns common in high-quality tips, such as emotional language or philosophical insights.
  5. Diversity-Weighted Re-ranking: To avoid showing the user ten tips that all say the same thing, the authors use a Determinantal Point Process (DPP) combined with PLSA (Probabilistic Latent Semantic Analysis). This ensures the final selection covers different "topics" of the song (e.g., one tip about the lyrics, one about the singer’s voice, and one about the emotional atmosphere).

Datasets

The authors release MTIPS, the first annotated Chinese dataset for music tip generation.

  • Source: ~50,000 reviews from a major Chinese music platform.
  • Scale: 8,003 labeled sentences (3,062 tips and 4,941 non-tips) across 128 songs.
  • Diversity: Covers five genres: Absolute music, Rock, Film & TV hits, Chinese pop, and EAJK (Europe, America, Japan, Korea) pop.
  • Annotation: Expert-labeled based on eight characteristics, including sensory artistic conception, humorous description, and background knowledge.

Key Insights and Ideas

  • Style is King: In the music domain, how something is said is almost as important as what is said. The ablation studies showed that removing the "Stylistic-based ranking" caused the most significant performance drop (nearly 29% in some metrics), proving that music tips require a specific "flavor" that standard summarizers lack.
  • The Popularity Proxy: By integrating "approval counts" (likes) into the TextRank algorithm, the model successfully identifies "representative" sentences that the community has already validated as high-quality.
  • Handling the "Comma Problem": The paper highlights a practical engineering hurdle: users rarely use periods. Their BERT-based punctuation restorer was essential for turning a messy stream of consciousness into extractable tips.

Key Results

  • Precision: GENTMS achieved a top-10 precision of 85.56% on seen songs, outperforming vanilla BERT by 3.34% and traditional models like SVM or FastText by over 10%.
  • Generalization: On "unseen" songs (songs not in the training set), the model maintained a high precision of 78.89%, demonstrating its ability to handle new musical releases.
  • Diversity: The diversity re-ranking module significantly improved the "Distinct" metric, ensuring that the generated tips weren't redundant. Qualitative results showed GENTMS could simultaneously highlight a song's rhythm, its cultural context, and its emotional impact.

Limitations and Future Work

  • Language Constraint: The current dataset and framework are optimized for Chinese. The linguistic patterns of "tips" (like specific rhymes or metaphors) may differ significantly in other languages.
  • Subjectivity vs. Personalization: While the model identifies "general" high-quality tips, music taste is deeply personal. The authors acknowledge that the next step is personalized tip generation, where the system selects tips based on an individual user's past listening habits and preferences.
  • Abstractive Generation: The current model is extractive (it pulls existing sentences). Future work could explore abstractive generation, where the AI writes entirely new, poetic tips from scratch.

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 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 learning-to-generate 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{zang2023learning,
  author = {Jingya Zang and Cuiyun Gao and Yupan Chen and Ruifeng Xu and Lanjun Zhou and Xuan Wang},
  title = {{Learning to Generate Tips from Song Reviews}},
  booktitle = {Neural Networks},
  year = {2023},
  url = {https://openreview.net/forum?id=6pG0g9WhoC},
}

Loading related content...

Figure