Context-aware Japanese reading tagger outperforms dictionaries on homographs
A new open-source model, ja-reading-tagger-v1, assigns context-appropriate Japanese readings to text spans, producing both written readings for ruby display and pronunciation readings for speech synthesis. On a sealed test of 5,000 sentences it slightly edges out the Sudachi dictionary baseline overall, scoring 0.9354 against 0.9308. The model performs markedly better on a 1,000-sentence set built around context-dependent homographs, reaching 0.7230 versus 0.5570. However, on real prose alone the dictionary remains slightly stronger, and the developers have stated this trade-off openly. The model is released under Apache 2.0 with ONNX support and runs in about 18 milliseconds per sentence on CPU, but it excludes personal names.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in