Natural Language Understanding in the Music Domain: Architecture, Features, and Challenges

The presentation, based on Qin Bin's talk at the DataFunTalk AI salon, introduces the research background and objectives of applying natural language processing techniques to the music domain, highlighting the unique problems and challenges of this vertical.

The overall backend service architecture of the Xiao Ai voice interaction platform is described, showing how Xiaomi Brain serves as a platform that abstracts SDK interfaces for manufacturers, integrates ASR services from various providers, and routes transcribed text to an NLP module that selects appropriate domain knowledge bases.

Key functionalities for the music vertical are outlined, including personalized recommendation, search intent handling (extracting singer, song, album, tag slots), disambiguation of ambiguous queries, error correction for ASR and user utterances, context inheritance, sentiment analysis, playback order control, lyric‑based song identification, and historical listening queries.

Challenges specific to music NLU are discussed, such as the complexity and variability of entity names, massive and noisy knowledge bases, unstructured user utterances, and diverse error types (homophone, dialect, word order). To address these, a knowledge‑base‑plus‑search solution is employed, leveraging Lucene‑based indexing, learn‑to‑rank (LambdaMart) re‑ranking, and feature‑rich GBDT models that incorporate slot matching, similarity, document popularity, and user feedback.

The data pipeline is explained, covering data acquisition from partners and crawlers, normalization, deduplication, labeling, and index construction, noting the substantial effort required for cleaning and maintaining millions of music records.

Further components include a recommendation intent classifier using n‑gram features and logistic regression (later enhanced with Word2Vec), custom grammar rules for intent bias detection, and a two‑stage ranking process that selects the most relevant document from the top candidates.

Finally, the article outlines current issues such as over‑recall, slot extraction accuracy, incomplete knowledge bases, and proposes future directions like end‑to‑end click‑model training and leveraging historical query similarity for faster response.