New BART model offers low-latency spell correction for Japanese search queries

Researchers have developed a compact BART-based sequence-to-sequence model for low-latency spell correction of Japanese music search queries. The model addresses challenges posed by the coexistence of four writing scripts by employing a script-aware synthetic misspelling generation pipeline. This pipeline incorporates keyboard-layout models, phonetic confusion priors, and kana case errors, while normalizing mixed-script titles to a single canonical script. Experiments show the model achieves 41.09% exact-match accuracy and a 11.62% character error rate, outperforming baselines with sub-4ms inference latency on a single GPU. AI

IMPACT Improves search relevance and user experience for Japanese language queries, potentially influencing future NLP research in multilingual spell correction.

RANK_REASON The item is an academic paper detailing a new model and methodology for a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• BART
• CatalyzeX
• DagsHub
• Gotit.pub
• graphics processing unit
• Hugging Face
• Japanese
• QWERTY
• ScienceCast

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →