FINB: a Japanese named entity recognition model based on multi-feature integration method.

This article focuses on improving Japanese named entity recognition (NER) by proposing a model called FINB (Feature Integration Network with BERT) that integrates multiple linguistic features inherent to Japanese text. Unlike English, Japanese lacks clear word boundaries and uses three character types—hiragana, katakana, and kanji—posing challenges for NER. FINB extracts pronunciation features via romanization (Romaji) and glyph features by converting characters into images, combining these with character embeddings in a BERT-BiLSTM-CRF architecture to enhance semantic understanding. Experiments on the Kyoto University Web Document Leads Corpus (KWDLC) and Japanese Wikipedia datasets demonstrate that FINB outperforms existing models, effectively reducing errors caused by homophones and improving recognition accuracy across various entity types. The study suggests that multi-feature integration leveraging language-specific characteristics can advance NER performance in Japanese and potentially other languages with similar complexities.