Towards an Efficient Code-Mixed Grapheme-to-Phoneme Conversion in an Agglutinative Language: A Case Study on To-Korean Transliteration

Won-Ik Cho, Seok Min Kim, Nam Soo KimCodeSwitch@LREC 2020 · OpenReview

Introduction

Grapheme-to-Phoneme (G2P) conversion—the translation of written text into phonetic sequences—is a foundational component of modern Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) systems. While G2P for monolingual text is a well-established field, real-world communication is increasingly "code-mixed," blending multiple languages, scripts, and symbols. In the context of Korean, this involves the frequent integration of English acronyms, Chinese characters (Hanja), and numerical data.

The Agglutinative Challenge

Korean presents a unique challenge for code-mixed G2P due to its agglutinative nature. Functional particles, which denote grammatical roles like subjects or objects, are attached directly to the ends of words without spaces. When a foreign term is introduced, these particles attach to it, creating complex phonetic interactions at the boundary. Standard G2P systems often fail to accurately segment the foreign "core" from the Korean "suffix," leading to unnatural prosody or incorrect pronunciation.

A Multi-Stage Pipeline for Code-Mixed G2P

To address these complexities, researchers from Seoul National University proposed a modular "divide and conquer" pipeline. This approach treats transliteration as a pre-processing step, converting all non-Korean elements into a uniform Hangul representation before applying final phonetic rules.

1. Detection and Morphological Analysis

The system first identifies "eojeols" (whitespace-separated blocks) containing non-Korean characters using Unicode properties. A morphological analyzer (MeCab-ko) is then employed to separate the foreign term from any attached Korean functional particles.

2. Hybrid Transliteration

Converting foreign words into Hangul is achieved through a two-pronged strategy:

  • Dictionary Lookup: A codebook of approximately 37,000 common English-Korean pronunciation pairs provides high-fidelity results for frequent vocabulary.
  • Neural Seq2Seq: For out-of-vocabulary terms, an Attention-based Recurrent Neural Network (RNN) predicts character-level transliterations.

3. Specialized Mapping and Contextual Logic

The pipeline includes specific modules for different data types:

  • Hanja (Chinese Characters): These are directly mapped to their established Sino-Korean Hangul equivalents.
  • Numbers: Context-aware logic determines whether a number should be read in English, Sino-Korean, or native Korean based on the surrounding linguistic environment.
  • Acronyms: The system distinguishes between strings that should be read letter-by-letter and those that function as whole words.
Examples of script-to-phoneme mapping for Hanja and Hangul.

Key Insights: Hangul as a Phonetic Pivot

The core innovation of this research is the use of Hangul as an intermediate phonetic representation. Because Hangul is a featural writing system—where the visual structure of characters reflects phonetic properties—converting foreign text into Hangul allows the system to leverage existing, highly accurate Korean G2P libraries (such as KoG2P or g2pK) for the final output. This "pivot" strategy ensures that the complex phonetic interactions between a transliterated foreign word and its attached Korean particles are handled correctly by the final G2P stage.

Results and Efficiency

The system demonstrates high efficiency by avoiding the computational overhead of massive end-to-end multi-lingual transformers. By modularizing the handling of Hanja, Latin script, and numbers, the pipeline remains lightweight and suitable for practical integration into real-time TTS systems. Case studies show successful decomposition and recombination of complex strings, such as correctly processing technical terms like "G2P" within a Korean sentence structure.

Future Directions

While the current framework is robust, future work aims to improve the handling of proper nouns and brand names, which often deviate from standard transliteration rules. The researchers also suggest that incorporating an external memory structure could help the system adapt to the rapidly evolving technical jargon and slang found in modern digital media. This multi-stage "detect-transliterate-G2P" framework holds promise for other agglutinative languages, such as Japanese or Turkish, that frequently feature mixed-script text.

Summary updated

BibTeX
@inproceedings{cho2020towards,
  author = {Won-Ik Cho and Seok Min Kim and Nam Soo Kim},
  title = {{Towards an Efficient Code-Mixed Grapheme-to-Phoneme Conversion in an Agglutinative Language: A Case Study on To-Korean Transliteration}},
  booktitle = {CodeSwitch@LREC},
  year = {2020},
  url = {https://openreview.net/forum?id=8A7D3K0lm5},
}
Figure