When a system writes an English name in Thai or Chinese script, there are usually several acceptable spellings. When it writes a Thai or Chinese name back in English, there is often one conventional spelling. The NEWS 2018 shared task found that this reverse direction, called back-transliteration, beats forward transliteration by roughly 15% F-score (Report of NEWS 2018 Named Entity Transliteration Shared Task). Synthetic data also helps one direction and hurts the other (Neural Machine Translation Techniques for Named Entity Transliteration). The field records this asymmetry but has not explained it. This note borrows an explanation from motor control.
In motor control, the degrees of freedom problem (Bernstein's problem) describes a body with more joints and muscles than any task needs. Many different movements reach the same goal. The nervous system settles on one movement by satisfying the task constraint and then applying a cost, such as smoothness or effort, to choose among the movements that remain. We propose to treat forward transliteration the same way. The task constraint is that a reader of the target script can recover the source pronunciation. The leftover freedom is orthographic convention, meaning which spelling native writers actually use. The claim is that much of the forward penalty comes from measuring a redundant choice against a single reference, not from the model failing to carry the sound across. A model that enforces the constraint and the cost separately should close more of the gap than one that learns a blurred average of both.
Core Claim
We make two claims, each of which can fail. The first is diagnostic. If outputs are judged against the full set of spellings native speakers accept, the forward/back gap should shrink in proportion to how redundant the target script is for that language pair. Redundancy here means the size or entropy of the acceptable set.
The second is methodological. A decoder that keeps only candidates whose pronunciation can be recovered, and then picks among them with an explicit cost learned from native names, should improve forward exact-match accuracy most on high-redundancy pairs. The cost plays the role of minimum torque-change or effort. A plain round-trip reranker should do worse on those same pairs, because recoverability alone does not decide which of the acceptable spellings natives prefer.
If the gap persists under set-based judgment, the asymmetry is real information loss rather than redundancy. That would be a useful negative finding.
Why the motor theory is a lens here, not a module
The obvious way to join these directions is to insert a simulated vocal tract between scripts. That would mean passing text through articulatory primitives such as the eight primitives of Motor control primitives arising from a learned dynamical systems model of speech articulation, or a Bayesian effort-weighted inversion as in Bayesian Modeling in Speech Motor Control. We reject this design, for two reasons.
First, those models were built on synthetic vowel-consonant-vowel sequences in a single language. Second, and more important, they cannot restore information the script never wrote down. Tamil does not distinguish voiced from unvoiced stops (A Large-scale Evaluation of Neural Machine Transliteration for Indic Languages), and Hebrew omits short vowels. No vocal tract physics recovers which voicing or which vowel was intended. A reviewer would also point out, correctly, that discrete phonetic features would carry whatever articulatory signal is useful, at a fraction of the cost.
What does transfer from motor control is the logical structure: an underdetermined mapping, resolved by a constraint plus a cost.
- A Computational Model for Cursive Handwriting Based on the Minimization Principle and Simulation of Optimal Movements Using the Minimum-Muscle-Tension-Change Model show that a single explicit cost picks one movement out of a redundant set.
- The Bayesian speech model shows that the weight on effort trades off against precision, depending on conditions.
- The volitional control of individual motor units is constrained within low-dimensional neural manifolds by common inputs supports the broader premise that redundant systems vary mainly along a few task-irrelevant directions.
These papers motivate the design. They supply no code we reuse.
Research Program
Measuring redundancy. We first measure redundancy directly. For each language pair available in both directions, we take a sample of test names (English–Thai and English–Chinese from NEWS 2018, English–Hindi and English–Tamil from Aksharantar, English–Hebrew from the Goldwasser et al. set). For each name we pool the references, the 10-best outputs of several systems, and outputs from the strongest model. Two native annotators then mark every candidate as acceptable or not, and mark the one they would write themselves. From these judgments we compute per name:
- the number of acceptable spellings;
- the entropy of the annotators' preferred spellings;
- a recoverability score: whether a back-transliteration model maps the candidate to the original source.
This gives, for each direction, an estimate of how much freedom is irrelevant to the task. It also lets us test whether the forward penalty tracks that freedom across pairs.
Decoding with a constraint and a cost. Given a base forward model $p(y\mid x)$, where $x$ is the source name and $y$ a target candidate, we score each candidate in its n-best list:
$$ s(y) = \log p(y \mid x) + \alpha \,\log p_{\text{back}}(x \mid y) + \beta \,\log p_{\text{conv}}(y) $$
The three terms are:
- $\log p(y \mid x)$: the base model's own score.
- $p_{\text{back}}$: a separately trained back-transliteration model. This term carries the task constraint that the pronunciation be recoverable.
- $p_{\text{conv}}$: a character language model trained only on native-script names and loanwords. This term carries the convention cost.
The weights $\alpha$ and $\beta$ are tuned on development data, separately for each language pair. The analogy to motor control predicts a pattern in these weights: $\beta$, the convention weight, should matter most where measured redundancy is high, and $\alpha$ most where the source script underspecifies the sound.
Training with a set-valued target. As a second stage, we fine-tune the base model with a minimum-risk objective. The reward is 1 for any spelling in the annotated acceptable set and 0 otherwise. This keeps the model from spreading probability across all valid spellings, which would leave the convention choice to chance.
Baselines
Naive reference. A single character-level attention sequence-to-sequence model trained on the same data, decoded greedily. It shows how large the forward/back gap is with no special handling. On English–Thai development data, Grundkiewicz and Heafield report 0.503 word accuracy for their baseline (Neural Machine Translation Techniques for Named Entity Transliteration).
Strongest specialized system. We re-implement the NEWS 2018 winning recipe: variational dropout, four-model ensembling, right-to-left rescoring, and back-translated synthetic data. It reaches 0.728 English–Thai development word accuracy in the same paper. It is the strongest competitor because it already reranks with a second model, so any gain from our approach must come from the specific constraint and cost, not from reranking in general. Our decoder runs on this system's n-best list with the same list size, the same training data, and the same development-only tuning budget.
Indic systems. For the Indic pairs, the strongest reported systems are:
- fine-tuned GPT-4o: 71.10% top-1 accuracy on Hindi AK-NEI, against 60.94% for IndicXlit;
- GPT-4.5: 78.48% on Tamil AK-Freq, against 69.75% for IndicXlit.
Both results are from Beyond Specialization: Benchmarking LLMs for Transliteration of Indian Languages. We cannot fine-tune GPT-4o under matched compute. We therefore compare to those published figures, and we also apply our reranker to the n-best outputs of an open model (Gemma-3-27B-it, which that paper also evaluates). This tests whether the method adds anything on top of a large language model.
Hebrew. The ILP aligner of Transliteration as Constrained Optimization reports an MRR of 0.894. That is a candidate-ranking task, so we compare on that MRR using their 300-pair test set.
What counts as beating a baseline. We need a statistically significant gain in forward exact-match accuracy over the reranked NEWS recipe on at least the high-redundancy pairs, and no loss in the back direction.
Ablations. The ablations isolate the key ingredient. We set $\beta=0$, which leaves only the round-trip constraint (essentially noisy-channel reranking). We set $\alpha=0$, which leaves only the convention cost. We also replace the native-name language model with a generic target-text language model. This tests whether convention specifically, rather than fluency in general, is doing the work.
Why not just…?
Why not just rerank by round-trip back-transliteration? This is the most serious threat, because it is our method with $\beta=0$. The redundancy argument predicts it will fall short exactly where many spellings recover the source equally well, since the constraint then fails to decide between them. If the $\beta=0$ ablation matches the full decoder on English–Thai and English–Chinese forward, our methodological claim fails. The diagnostic claim would still stand.
Why not just report multi-reference F-score, which already gives partial credit? Character F-score rewards near-misses whether or not they are acceptable spellings. It cannot tell "a valid alternative convention" apart from "a wrong sound". Moreover, NEWS reference sets are incomplete for many names. Annotated acceptable sets are what let us separate redundancy from error. We will report F-score as well, so readers can see how far it diverges from set accuracy.
Why not just prompt or fine-tune a large language model? Large language models are strong on Indic benchmarks. However, the same paper documents that they sometimes translate the meaning of a word instead of transliterating its sound. Their gains also come with no account of why forward and back directions differ. Our reranker can be placed on top of their n-best lists. If it adds nothing there, that suggests large models already resolve convention internally, and we would report that.
Why not just use IPA or articulatory features as an intermediate step? This is a reasonable alternative to a vocal tract model, but it addresses the wrong side of the problem. A phonetic intermediate helps carry the sound across, which is the task constraint. It does not choose among target spellings that sound the same. Discrete intermediate pipelines are also known to miss target orthographic conventions (Report of NEWS 2012 Machine Transliteration Shared Task). We will include a phoneme-feature back model as a variant of $p_{\text{back}}$, to check whether it strengthens the constraint term.
Proposed Experiments and Decisive Checks
The decisive diagnostic. For each pair available in both directions, we compute the forward-minus-back gap twice: once under single-reference exact match, and once under annotated set accuracy. We then regress the shrinkage of the gap on measured redundancy (log acceptable-set size and preference entropy). We do this across English–Thai, English–Chinese, English–Hindi, English–Tamil and English–Hebrew.
- Redundancy hypothesis supported: a positive relationship, with set accuracy recovering most of the roughly 15% gap on high-redundancy pairs.
- Redundancy hypothesis refuted: the gap is unchanged under set accuracy. That would point to information loss as the real cause, which Tamil's underspecified stops would make plausible in the Tamil-to-English direction.
The decisive method check. We compare the ordering of the full decoder against its $\beta=0$ and $\alpha=0$ ablations, on forward exact match. This runs on NEWS 2018 English–Thai and English–Chinese, on Aksharantar Hindi AK-NEI and Tamil AK-Freq, and on the Hebrew set.
- Predicted: the convention term helps most on the pairs measured as most redundant, and the recoverability term helps most on Hebrew and Tamil.
- If the gains do not line up with measured redundancy, the motor-control analogy fails to predict anything, and we would say so.
Secondary checks. We test whether the tuned weights $\alpha$ and $\beta$ move with redundancy and underspecification as predicted. We also test whether training with the set-valued objective reduces the harm that forward synthetic data caused in the NEWS 2018 systems.
References
- NEWS 2018 Whitepaper
- Neural Machine Translation Techniques for Named Entity Transliteration
- Motor control primitives arising from a learned dynamical systems model of speech articulation
- Bayesian Modeling in Speech Motor Control: A Principled Structure for the Integration of Various Constraints
- A Large-scale Evaluation of Neural Machine Transliteration for Indic Languages
- A Computational Model for Cursive Handwriting Based on the Minimization Principle
- Simulation of Optimal Movements Using the Minimum-Muscle-Tension-Change Model
- The Volitional Control of Individual Motor Units Is Constrained within Low-Dimensional Neural Manifolds by Common Inputs
- Beyond Specialization: Benchmarking LLMs for Transliteration of Indian Languages
- Transliteration as Constrained Optimization
- Report of NEWS 2012 Machine Transliteration Shared Task