Letting Syllable Structure Decide When to Translate and When to Transliterate

Translating a name across writing systems forces a choice for each piece of it. Some pieces should be carried over by sound, which is transliteration ("Delhi" becomes 德里). Others should be carried over by meaning, which is translation ("New" becomes 新). Real names mix the two: "New Delhi" becomes 新德里, and "Ministry of Health" or "Faeroe Islands" work the same way. For languages with only a few hundred to a few thousand example pairs, the best existing systems are small statistical models. These models learn sound correspondences between character strings, and they depend on hand-written rules to keep meaning-bearing words out of those sound tables.

This proposal argues that a constraint known from phonology can do that sorting automatically. The constraint is that a transliteration must be built from legal syllables of the language, and that each source syllable maps to a target sound monotonically and locally. A transliteration model restricted to this syllable structure has too little capacity to explain a semantic pair such as "Ministry" and its native-language equivalent. If such a model competes with a word-translation table inside one probabilistic model, the data should route each word, or each run of syllables, to the right channel on its own. The two-sentence version: put a syllable-constrained sound model and a meaning model in competition inside one latent-variable aligner. The phonotactic constraint then becomes the signal that separates transliteration from translation, with no stop-lists, classifiers, or dictionaries required.

Core Claim

The closest prior system is Learning Better Name Translation for Cross-Lingual Wikification. It learns name translation from noisy Wikipedia title pairs and treats word alignment as a hidden variable. Its transliteration component works on flat character strings. Its authors still had to exclude the ten most frequent words from transliteration training by hand, because otherwise words like "Department" polluted the character tables. We read that heuristic as a symptom. A flat character model is flexible enough to memorize a semantic pair as a strange multi-character "sound" mapping. Expectation-maximization (EM) therefore has no reason to prefer the translation explanation.

Our claim is that restricting the sound channel to syllable-structured, phonotactically legal outputs makes the choice between translation and transliteration identifiable from the data alone. This should also improve transliteration accuracy and syllable validity in the low-resource regime. The second half of the claim is supported by Phonology-augmented statistical transliteration for low-resource languages, whose onset–nucleus–coda pseudo-syllable model works with as few as 94 Vietnamese word pairs. That system handles only pure transliterations of single words. The first half of the claim, that syllable structure acts as a routing regularizer, has not been tested anywhere in the material we found.

The Model

A source name $s = s_1 \dots s_n$ is paired with a target name. The target is segmented into syllables $t = t_1 \dots t_m$. A hidden alignment $a$ assigns each source word $s_i$ a contiguous span of target syllables $t_{a_i}$. A hidden route $z_i$ marks that span as translated or transliterated:

$$p(t \mid s) = \sum_{a,z} \prod_{i=1}^{n} \Big[ \lambda_{s_i}\, p_{\text{tr}}(t_{a_i} \mid s_i) \Big]^{[z_i = \text{tr}]} \Big[ (1-\lambda_{s_i})\, p_{\text{syl}}(t_{a_i} \mid s_i) \Big]^{[z_i = \text{tl}]}$$

The terms are defined as follows. $p_{\text{tr}}$ is a word-to-span translation table. $p_{\text{syl}}$ is a monotone transducer that maps learned source substrings onto target syllables. $\lambda_{s_i}$ is a learned, word-specific prior on translating. The indicator $[z_i = \cdot]$ selects the channel. Both the alignment and the route are summed out with EM.

Two design choices carry the claim. First, $p_{\text{syl}}$ can only emit syllables that occur in a target-side syllable inventory, and it is scored by an onset–nucleus–coda syllable language model in the style of Ngo et al. Its per-syllable table is small and shared across the whole corpus. Second, the route operates on spans of syllables rather than whole tokens. This lets 新德里 be split into one translated syllable and two transliterated ones, a case the hard pre-classifier of Lightly Supervised Transliteration for Machine Translation cannot represent.

Target-side syllabification is nearly free for many of the scripts that matter. Vietnamese writes syllables separated by spaces. Each Han character is one syllable. Brahmi-derived scripts segment into aksharas by Unicode rules, which A More Discerning and Adaptable Multilingual Transliteration Mechanism for Indian Languages argues are the natural matching unit. We do not syllabify the source side by rule. The rule-based English syllabifiers in Syllable-Based Machine Transliteration with Extra Phrase Features and A Hybrid Transliteration Model for Chinese/English Named Entities are a known weakness on loanwords. Instead, source substrings are learned as part of the alignment, following the substring transducer of Substring-Based Transliteration. That paper also shows that substring units compress the model drastically (759 versus more than 86,000 transducer states). The same compression is what lets multi-word entities avoid the length and distortion problems that Character-Based PSMT for Closely Related Languages ran into with flat characters.

For directions whose target is English, such as Bengali→English in the wikification setting, the syllable constraint is placed on the source side instead. Source aksharas are the units, and each must map to an English substring under a monotone, length-bounded transducer. We expect the capacity argument to hold there as well, but more weakly, and the experiments test this separately.

Research Program

The work has three parts.

The first part builds the joint model. It is initialized from a pure-transliteration seed, which can come from NEWS pairs where available or from a few hundred pairs otherwise. It is then trained on Wikipedia title pairs subsampled to 100, 300, 1,000, and 3,000 pairs. This range covers the regime where Ngo et al. found that unconstrained statistical models needed more than 1,500 pairs to catch up.

The second part is an annotated evaluation set. For each language, a few hundred held-out hybrid entity titles (organizations, locations, administrative units) will be marked word by word, or syllable span by syllable span, as translated or transliterated. Routing accuracy can then be measured directly instead of being inferred from downstream scores. No such gold annotation exists in the material we found, so this is a real contribution of the project.

The third part is an analysis of why routing works or fails. We will measure how sharply the posterior over $z_i$ separates frequent semantic words from names as the size of the syllable inventory and the strength of the phonotactic model are varied. This tests the capacity explanation directly rather than relying on end accuracy.

Languages are chosen to vary script type. Bengali and Hindi represent Brahmi scripts. Vietnamese represents a tonal language with space-separated syllables, and connects to the IARPA Babel data used by Ngo et al. Chinese represents a logographic script with sub-token hybrids. Tagalog is a Latin-script control where the syllable constraint should matter least. We also add one or two long-tail languages from the 292 covered by the Illinois Cross-Lingual Wikifier that have few title pairs.

Baselines

The naive reference point is exact-match lookup in Wikipedia inter-language links, plus identity copying for Latin scripts. Learning Better Name Translation for Cross-Lingual Wikification reports that this lookup reaches a Bengali candidate coverage upper bound of 20.82%. Any learned system must clear this bar by a wide margin.

The strongest existing method for the question is the joint latent-alignment model of Tsai and Roth 2018, with its top-ten stop-list and flat character transliteration. It is the only prior system that handles mixed translation and transliteration from noisy title pairs in low-resource languages. It reports Bengali candidate coverage of 65.18%. We will reimplement it on identical title-pair subsamples, give it the same number of EM iterations and the same development-set tuning budget for its smoothing and its stop-list size, and report coverage under its own protocol. Beating it means higher candidate coverage on Bengali at every data size, with the clearest margin at or below 1,000 pairs. It also means higher top-1 accuracy on the annotated hybrid subset, with no stop-list in our model.

The key ablation removes the syllable constraint while keeping everything else identical, including the learned prior $\lambda$ and the span-level routes. The sound channel becomes a flat character transducer, and the stop-list is also removed. If this ablation routes as well as the full model, the central claim is false.

A second ablation keeps the syllable channel but routes whole tokens rather than syllable spans. This isolates the value of sub-token routing on Chinese and Vietnamese hybrids.

For pure-transliteration quality, we include the reranked Moses plus Sequitur system from Statistical Machine Transliteration Baselines for NEWS 2018. For the neural comparison, we include the deep GRU ensemble with right-to-left rescoring and back-translated synthetic data from Neural Machine Translation Techniques for Named Entity Transliteration, which reports English–Thai word accuracy of 0.728. Both are trained on the same subsamples. At full NEWS scale we do not expect to beat the neural system. The comparison exists to locate the data size at which its advantage appears.

Why not just…?

Why not just drop the stop-list from Tsai and Roth and let EM learn a per-word translation prior? This is the most serious alternative, and it is exactly the flat-character ablation above. Our argument is that a flat character channel can absorb semantic pairs as long multi-character mappings, so a learned prior has nothing pushing it toward the translation route. The fact that Tsai and Roth needed a manual cutoff at all is indirect evidence for this. Indirect evidence is not proof, though. If the ablation matches the full model on routing accuracy, we will report that syllable structure only helps output validity and not routing, and narrow the claim to that.

Why not just add a syllable-validity filter on top of an existing system's outputs? A filter in the style of Ngo et al.'s Invalid Syllable Rate check would remove illegal outputs cheaply. It acts only after training, though. By then, semantic words have already contaminated the transliteration table and shifted the alignments of genuine names. We test this by reranking the Tsai and Roth reimplementation with our syllable language model. If that closes most of the gap in coverage and hybrid accuracy, the training-time constraint is not justified.

Why not just use character-level neural models with synthetic data, or prompt a multilingual language model? Tsai and Roth report that character neural translation falls short on Bengali and Tagalog. Grundkiewicz and Heafield show that back-translated synthetic data helps substantially at NEWS scale. We found no reported result on Lacuna for few-shot large-language-model transliteration of hybrid entities. This is a genuine threat for well-represented languages like Hindi or Chinese. We will run a prompted multilingual model with the same in-context examples drawn from each subsample, and report results separately for languages inside and outside its likely training coverage. If it dominates everywhere, the contribution shrinks to truly long-tail languages and to interpretable routing.

Why not just run a classifier first and send each word to the right model? The pipeline in Lightly Supervised Transliteration for Machine Translation reaches 90.26% identification accuracy. It cannot recover from its own errors, and it cannot split a single token like 新德里. It also needs part-of-speech resources that low-resource languages lack. Our span-level routing evaluation on Chinese and Vietnamese hybrids measures this limitation directly.

Decisive Checks

The first and most important check is routing accuracy on the annotated hybrid sets, comparing the full model against the flat-character ablation without a stop-list at 100 to 3,000 title pairs. The claim stands only if the syllable model separates translated from transliterated spans clearly better than the ablation, and if that margin is largest at the smallest data sizes.

The second check is candidate coverage on Bengali against the reimplemented 65.18% baseline, together with end-to-end F1 on the TAC 2016 entity linking evaluation, where Tsai and Roth report Chinese F1 rising from 63.62 to 65.53.

The third check is Invalid Syllable Rate and token error rate on the Babel Vietnamese partitions, to confirm that adding the translation channel does not erode the gains Ngo et al. obtained from pure syllable modeling.

The fourth check is the Tagalog control. If the syllable constraint helps there as much as it helps in Bengali or Chinese, the effect is probably coming from general regularization rather than from script-specific phonotactics. We would report that explanation instead.