- The paper introduces Reddit2Deezer, a scalable dataset that links Reddit music discussions with Deezer’s catalog to support robust conversational recommendation research.
- The paper employs a two-stage LLM filtering and deduplication pipeline, achieving an average Deezer mapping accuracy of 93.84% for reliable data grounding.
- The paper benchmarks CLAP-based retrieval and LLM-based generative models, highlighting improved performance with paraphrased dialogues and multi-turn context.
Overview of Reddit2Deezer: A Reality-Grounded Corpus for Conversational Music Recommendation
The "Reddit2Deezer" dataset (2605.09120) addresses longstanding challenges in the study of conversational music recommendation (CMR), centering on the limited scale and rapidly degrading grounding of authentic dialogue resources. By systematically mining Reddit’s music-related communities and linking conversation-sourced recommendations to Deezer identifiers, this resource supports both scalable research and robust content-based modeling for CMR, while emphasizing the reproducibility and versatility of its data formats.
Motivation and Context
Prevailing work in CMR has faced a core tradeoff: hand-collected, authentic conversation datasets are small in scale, and large dialog corpora generated via LLM or simulated interaction pipelines lack the distributional fidelity of real user behavior. Simultaneously, many existing datasets are hampered by reliance on platforms (e.g., YouTube, LFM-2b, or proprietary logs) that degrade over time or are no longer accessible, restricting both catalog coverage and independent reproducibility. The authors position Reddit2Deezer as addressing these deficiencies: it is built from real user conversations from 200 subreddits, filtered and annotated using high-confidence LLM pipelines, and linked exhaustively to Deezer’s robust, publicly accessible catalog—supporting downstream access to audio, metadata, and longer-term curation.
Corpus Construction Pipeline
The curation strategy for Reddit2Deezer is distinguished by several methodological strengths:
- Community-Sourced Coverage: The dataset begins with a comprehensive registry of 695 Reddit music subreddits, narrowed by manual topical filtering and API-based accessibility checks, yielding a structurally broad, diverse source pool.
- Hybrid Filtering Protocol: After structural and content-level prefilters to prune non-music-recommendation threads, a two-stage LLM (Qwen3.6-35B-A3B-FP8) filter verifies recommendation intent and extracts explicit artist-title information, enforced by a strict confidence threshold.
- Deduplication and Traceability: The pipeline systematically addresses redundant cross-posting within and across subreddits, collapsing duplicates via normalized titles and retaining post IDs for traceability.
- Deezer Grounding: Recommendations are linked via exact artist-title matching, securing stable access to audio previews and metadata beyond the volatility of YouTube or Spotify URIs.
- Paraphrased and Raw Releases: To reconcile authenticity and reproducibility, two dataset versions are released. The paraphrased variant, generated by a dedicated LLM prompt, restyles exchanges as canonical one-on-one music recommendation chats, removing Reddit-specific artifacts and personal information, while preserving all music-relevant details.
Data Validation and Statistical Highlights
A human validation study, guided by stringent sample size estimation and dual annotator protocols, confirms that the majority of conversations are genuine music recommendation exchanges and that Deezer mappings are robustly accurate (average correctness of 93.84%). The paraphrased version achieves extremely high faithfulness ratings (mean ∼4.8/5), while also nearly tripling conversation count due to the one-to-one mapping per {thread, comment, item} triple.
Key corpus statistics include approximately 186k raw and 536k paraphrased conversations, with coverage of over 100k tracks and 29k albums. Unlike nearly all previous CMR datasets, album-level recommendations are systematically retained and linked. The corpus is predominantly single-turn, though it includes a nontrivial quantity (∼2k) of multi-turn dialogs.
Experimental Evaluation and Comparative Results
The study benchmarks two model classes for conversational music recommendation on both corpus variants: CLAP-based retrieval models (leveraging contrastive joint audio-text encoders) and LLM-based generative models (Qwen3.5-2B, both zero-shot and fine-tuned variants). Evaluation metrics include Hit@k, Recall@k, and nDCG@k.
Principal findings:
- CLAP-Audio outperforms CLAP-Text in nearly all metrics, establishing the superiority of audio-based semantic retrieval over text-centric approaches when matching seeker queries to catalog items.
- Both retrieval models perform better on the paraphrased test set, reflecting improved domain alignment and reduced platform-specific noise.
- Zero-shot Qwen3.5-2B generative modeling substantially surpasses the retrieval variants, confirming that LLMs encode substantial music-relevant world knowledge for this task even without in-domain training.
- Fine-tuned LLMs (FT-Raw, FT-Para) exhibit expected domain adaptation: FT-Raw is strongest on the authentic Reddit-styled test partition, while FT-Para is strongest on the paraphrased partition. Notably, FT-Para's training on cleaned, canonical dialogues provides more transferable representations for cross-partition evaluation.
- Despite the dataset’s single-turn dominance, models improve their recommendations with longer dialogue context in multi-turn settings, exploiting the additional user preference signal.
Theoretical and Practical Implications
Reddit2Deezer brings several advances to the CMR domain:
- Enhanced Realism at Scale: By scaling up the extraction of real conversational behavior without resorting to full synthetic generation, this resource enables the training and evaluation of next-generation CMR systems on empirical user language and intent distributions.
- Robust Grounding and Open Accessibility: Leveraging Deezer as the canonical entity linker provides both direct audio/metadata access (without API key restrictions) and future-proofing against link decay. This allows for reproducible multimedia experiments and content-based modeling strategies.
- Novel Evaluation Opportunities: The dataset’s structure supports single- and multi-turn modeling, track/album granularity, and transfer experiments between authentic and standardized phrasings—facilitating research into domain adaptation and dialogue style normalization.
- Implications for LLM-based Recommendation: The strong results of relatively compact LLMs—both zero-shot and fine-tuned—suggest further opportunities for leveraging pretrained LLMs in music understanding and recommendation tasks, possibly in conjunction with audio-based representation learning.
Future Directions
Immediate avenues for future progress include extending multi-turn dialogue coverage, exploring richer user preference elicitation in conversation, and more tightly integrating content-based and collaborative filtering signals leveraging the dataset’s Deezer API connectivity. Further, Reddit2Deezer’s methodological framework could be adapted to construct similarly grounded conversation datasets for other domains (e.g., films, books, games) using analogous subreddit communities and robust entity linking.
Conclusion
Reddit2Deezer (2605.09120) establishes a scalable and reality-grounded standard for CMR research, blending the linguistic diversity of Reddit with stable, future-proof audio and metadata grounded in Deezer. Through comprehensive construction, validation, and benchmarking, the dataset is positioned to support rigorous, reproducible research in conversational music recommendation, while also informing the future integration of multimodal, LLM, and content-based approaches in this domain.