Music Domain Named Entity Recognition: Challenges, Solutions, and Future Directions

The talk introduces Named Entity Recognition (NER) as a core NLP task and focuses on its application in the music domain, where entities include song names, artist names, shows, genres, and versions.

Background and Development NER evolved from rule‑based and statistical methods (HMM, CRF, SVM) to shallow neural networks (LSTM/IDCN + CRF) and, after 2018, to large pre‑trained language models such as BERT, with recent research emphasizing transformer‑based approaches.

Domain‑Specific Challenges Music NER suffers from strong domain relevance, high name ambiguity, insufficient context in user queries, and diverse expression styles in music texts.

Overall Solution The system is split into offline and online modules. Offline processing builds basic data (search/play logs, music library), intermediate data (entity knowledge base, rule base, candidate sets, training corpora), and models. Online prediction first generates candidates, then applies rule‑based filtering and model inference, finally fusing results.

Candidate Generation & Training Data Candidates are extracted via a path‑selection algorithm that combines confidence, language‑model scores, and a Root‑Link metric, using beam search for efficiency. Training data is iteratively improved through active learning, weak supervision, and data augmentation (entity replacement, non‑entity replacement, name perturbation).

Model Iterations V1 uses handcrafted features with a GRU classifier for short queries. V2 treats NER as sequence labeling, employing BiLSTM‑CRF with domain‑fusion features. V3 adds feature self‑attention and multi‑view attention to better handle ambiguity and long‑range dependencies, achieving higher precision and recall.

Knowledge Fusion Frameworks Three families are discussed: Lattice LSTM (modifies RNN cells), CGN (adds graph neural networks), and FLAT (injects multi‑granularity token/word embeddings into transformer self‑attention). FLAT is adopted and further refined by adjusting positional encoding and incorporating pairwise relations.

Future Outlook Planned improvements include integrating knowledge‑graph representations for entities, jointly optimizing candidate generation with final prediction, and extending the model to support nested and discontinuous NER.