📄 A Joint Source-Channel Model for Machine Transliteration

Summary

The paper addresses machine transliteration of foreign personal names between English and Chinese, in both directions: forward (E2C) and back-transliteration (C2E). Earlier systems usually rely on a phoneme-based noisy channel model (NCM). That pipeline needs grapheme-to-phoneme conversion specific to each name's origin, followed by phonemes to Pinyin and then Pinyin to characters, and every stage adds errors. The authors propose direct orthographic mapping (DOM), which skips the phonemic layer. They realise it through a joint source-channel model, the n-gram transliteration model (TM). This model scores sequences of aligned English–Chinese unit pairs jointly rather than modelling one side conditioned on the other. Unit alignments are learned automatically from a bilingual name dictionary with EM, so no hand alignment is needed. The authors also implement NCM and an ID3 decision tree under DOM for comparison. On an English–Chinese name dataset, the 3-gram TM outperforms ID3 and achieves far lower error than previously reported phoneme-based results. This shows that orthographic context alone can drive accurate transliteration while reducing development effort.

Key points

• The data come from an English–Chinese personal name dictionary, with 34,777 training entries and 2,896 test entries.
• The 3-gram TM reaches a 10.8% character error rate in E2C, compared with 15.6% for ID3 under DOM.
• The best phoneme-based baseline is the "Huge MT (Big MT)" setup from Virga & Khudanpur (2003), with 42.5% Pinyin error. Against it, the paper claims a 74.6% relative error reduction, although the baseline was evaluated on different data.
• Word error rates for the 3-gram TM versus ID3:
  • E2C closed test: 1.6% vs 9.7%.
  • E2C open test: 29.9% vs 39.1%.
  • C2E closed test: 14.7% vs 38.4%.
  • C2E open test: 62.1% vs 63.3%.
  • C2E remains much harder in both settings.
• The motivation for avoiding Pinyin comes from Virga & Khudanpur, who found that converting Pinyin to characters alone costs 8.3% absolute accuracy because of homophones. Among the 374 characters used for transliteration, each character maps to a unique Pinyin, but one Pinyin can map to several characters.

Interesting ideas

• I1. Joint modelling of paired units instead of a noisy channel: Modelling the probability of the pair sequence directly yields a single model that can be marginalised for either direction. This could suit any bidirectional string-to-string task, such as Ukrainian–Latin romanisation and its reversal.
• I2. EM-induced alignment as a by-product: Starting from a random segmentation, the method alternates between re-estimating n-gram statistics and re-aligning the entries until the alignment converges. The final alignments then form the unit table used for decoding. This removes manual alignment and could bootstrap lexicons for low-resource name pairs.
• I3. Skipping intermediate representations: Mapping orthography directly avoids error compounding across pipeline stages and the need for G2P tools for each language of origin. This is attractive when names come from many source languages.
• I4. Variable-length units beat fixed windows: English units in the learned table span 1 to 7 letters. The authors attribute part of ID3's weaker performance to its fixed context window, which argues for segment-level models when unit lengths vary widely.
• I5. Separate closed and open evaluation in both directions: Reporting closed and open tests for E2C and C2E exposes the asymmetry between the two directions and how poorly models generalise to unseen names. It is a cheap diagnostic worth reusing.