Browse Teach Curate

Chinese Polyphone Disambiguation with Pre-trained BERT Semantic Features

An end-to-end grapheme-to-phoneme framework that predicts the pronunciation of a Chinese polyphonic character directly from a raw character sequence. A pre-trained BERT extracts contextual semantic features that a neural classifier maps to pronunciation labels, using a separate non-shared output layer per polyphonic character so new characters can be added without retraining the whole model.

Data Text

Disambiguation of Chinese Polyphones in an End-to-End Framework with Semantic Features Extracted by

This 2019 Interspeech paper proposes an end-to-end framework for Chinese polyphone disambiguation -- choosing the correct pronunciation of a character that has several, the core problem in Mandarin t…