Abstract
Music recommendation systems often suggest individual tracks, yet the detailed critiques explaining these recommendations are typically buried within lengthy album reviews. This paper presents a natural language processing pipeline designed to bridge this gap by extracting song-specific comments from full-length album reviews to generate "fine-grained" summaries. Evaluation shows that these targeted summaries significantly improve the speed and quality of user decision-making compared to reading raw album reviews.
Introduction
While most music recommendation systems operate at the song level, high-quality written critiques are predominantly found at the album level on platforms like Amazon or Pitchfork. Song-specific comments on these platforms are often too brief (e.g., "Great beat!") to be informative. The challenge lies in the fact that mentions of a specific song are frequently fragmented throughout an album review; a reviewer might mention the vocals of a track in the first paragraph and its guitar riff in the final sentence. To provide a helpful recommendation, a system must identify these disparate mentions, group them logically, and present them as a coherent summary.
Methodology
The authors proposed a five-stage summarization pipeline to transform album-level critiques into song-level insights.
- Title Extraction: Using n-grams and a song title lexicon, the system identifies song mentions. It employs context to distinguish between a song title and identical words appearing in lyrics (e.g., distinguishing "Maxwell's Silver Hammer" as a title vs. a line in the song).
- Feature Extraction: Utilizing "text tiling," the system divides reviews into sub-topic segments. It identifies "feature terms" (e.g., lyrics, riff, bass-line) using WordNet to pinpoint text describing the song.
- Sentence Partitioning and Regeneration: To handle sentences discussing multiple features or songs, the system uses the Stanford Parser for decomposition. A template-based realizer (YAG) then regenerates these fragments into independent, grammatically correct "f-sentences." These atomic units of opinion are easier to summarize than complex, multi-topic original sentences.
- Grouping: These f-sentences are categorized by sub-feature and sentiment polarity (positive vs. negative) using SentiWordNet.
- Selection and Ordering: The system selects representative sentences and orders them using a music ontology (e.g., Song → Sound → Instrument). This ensures the summary flows logically, starting with general impressions before moving into specific instrumental critiques, avoiding the "choppy" feel of many extractive summarizers.
Datasets and Evaluation
The primary corpus consisted of 1,350 album reviews from Amazon and Epinions, covering 27 albums. For evaluation, 100 reviews were manually annotated for song titles and features. Additionally, professional reviews from sites like Pitchfork and Rolling Stone served as a "gold standard" for informativeness, allowing the authors to measure how well the generated summaries captured "essential" song features.
Key Results and Insights
The system demonstrated high performance in extraction tasks, achieving 91.3% precision and 98% recall for title extraction, and 95.4% precision for feature extraction.
A user study involving the "SongRecommend" interface revealed several key insights:
- Efficiency: Participants using the summary-based interface made decisions significantly faster than those reading full album reviews.
- Informativeness: Automated summaries captured over 91% of the features identified in professional reviews, suggesting that crowdsourced reviews contain technical insights comparable to professional critiques when properly extracted.
- Reduced Choice Paralysis: While users reading full reviews often defaulted to the first recommendation, those given song-specific summaries explored a wider variety of options, indicating better differentiation between tracks based on the provided summaries.
Limitations and Future Work
Despite its success, the system does not currently perform "text cleaning," meaning spelling errors or informal slang from original reviews can persist in summaries. Handling conflicting opinions between reviewers—where one loves a riff and another hates it—also remains a challenge for generating a unified voice. Furthermore, the reliance on a domain-specific music ontology suggests that adapting this approach to other fields, such as movie or restaurant reviews, would require significant manual effort to define feature hierarchies and priority orders.
Summary updated