Running a low-resource language through a commercial large language model is expensive. The model's tokenizer, the component that cuts text into billable pieces, shreds native words into many fragments. The standard fix is to translate the text into English, or rewrite it in Latin script, before calling the model. Cost-Performance Optimization for Processing Low-Resource Language Tasks Using Commercial LLMs shows this cuts cost by about 90% and even raises accuracy. That result comes from benchmarks such as sentiment, summarization and math, whose answers rarely depend on local culture. Ethnographic work on intangible cultural heritage describes the same move in commercial settings. Sichuan opera face-changing becomes a tableside stunt in Challenges and Opportunities of Sichuan Opera Face Changing in Haidilao in Cultural Communication, and Yi folk songs lose their linguistic integrity on tourist stages in An Investigation of Yi Folk Song Knowledge and Development in Guizhou, China. In both cases, making heritage cheap to consume strips out the hard-to-carry local parts. This note asks the computational version of the question: when a cost-saving pivot runs over culturally dense text, exactly what is lost, and how much of it can be bought back, at what price?
Our central bet is that the loss is concentrated. Most of the cultural damage done by an English pivot falls on a small fraction of spans: kinship terms, ritual and festival names, foods, idioms, and forms of address. Each of these either becomes a generic English word ("a festival", "a sweet") or disappears. If that is true, the tradeoff is not smooth. Keeping only those spans in their native form, and translating everything else, should recover most of the cultural accuracy for a small fraction of the native-script cost. We propose to measure this frontier on literary text in real low-resource languages, and to build and test that selective pivot.
Core Claim
We define two quantities per document. $C$ is the number of tokens billed by the commercial model. $R$ is cultural retention: the fraction of annotated culture-bearing units whose meaning survives in the model's output, judged against native-speaker annotations. The first claim is diagnostic. Measured on culturally dense literary text rather than generic benchmarks, the English pivot falls well below native-script processing on $R$, even though Nag et al. find it matches or beats native processing on generic accuracy. The second claim is constructive. A pivot that protects detected culture-bearing spans reaches most of the native-script $R$ while its $C$ stays close to the plain pivot's. The frontier is sharply concave: the first few percent of protected tokens buy most of the retention.
This is a measurement and design question that none of the connected work poses. Nag et al. optimize cost against generic accuracy and never measure cultural content. They also find that word-by-word dictionary substitution breaks syntax. That finding constrains our design: we protect whole spans inside fluent translated sentences rather than swapping words into native syntax. MELLA: Bridging Linguistic Capability and Cultural Groundedness for Low-Resource Language MLLMs motivates the proposal. It shows that translated training data yields "fluent fools" that miss local connotation, with Czech cultural keyword accuracy at 1.10% before its native-alt-text tuning. MELLA addresses the problem at training time for national languages with large native websites. We address it at inference time, as a cost decision, for languages without that web presence. LittiChoQA: Literary Texts in Indic Languages Chosen for Question Answering supplies the test bed and a second compression axis. Its paragraph selection trades a small semantic-similarity drop for a large throughput gain, but it never checks what cultural framing is discarded.
Research Program
The first component is an evaluation set. It starts from LittiChoQA stories in Bhojpuri, Maithili and Awadhi, with Hindi as a mid-resource contrast. Native-speaker annotators mark culture-bearing units in each story. They then write questions whose correct answers depend on those units, for example who a kinship term refers to, what a ritual signifies, or what an idiom implies about a character. They also write a matched set of plot-factual questions that do not depend on those units. The matched set lets us separate cultural loss from general comprehension loss. LittiChoQA's own questions were generated by GPT-4.1, so we do not rely on them for the cultural half.
The second component is a detector for culture-bearing spans that needs no manual labels. It exploits the pivot's own failures. We translate native text to English with IndicTrans2 and back to the native language. A span counts as a candidate when its round trip fails to recover the source term, or when its English alignment is a generic hypernym. A second signal adds terms that are frequent in native literary text but rare in machine-translated text of the same language, which marks translationese. Mitigating Translationese in Low-resource Languages: The Storyboard Approach documents how translated text drifts from native lexis; we use that drift as a detection signal rather than trying to fix it by elicitation. A small tagger is trained on these weak labels and scored against the human annotations.
The third component is the protected pivot. Detected spans stay in native script, with a short English gloss in brackets, inside an otherwise English translation. The protected text goes to the commercial model under a fixed token budget. We sweep a threshold on the detector's confidence to trace the full $R$–$C$ curve. We cross this with the three context regimes from LittiChoQA, full story, paragraph selection and dense retrieval, to see whether shortening the context and pivoting the language compound each other's cultural losses.
Baselines
The naive reference point is native-script processing: the full story in native script, sent to the same commercial model with the same prompt. This is the retention ceiling and the cost ceiling. Nag et al. report that native text costs up to 10× more tokens than English.
The strongest existing method is the IndicTrans2 English pivot of Cost-Performance Optimization for Processing Low-Resource Language Tasks Using Commercial LLMs. It ranked best on their performance-to-cost ratio, with roughly 90% cost reduction and a 14.7% average sentiment-accuracy gain over native script. We also run their Latin transliteration strategy, which they report cuts tokens 2× to 7×.
For the context axis, the reference numbers are Krutrim-2 12B results from LittiChoQA on STS MuTe, a semantic-similarity score for answers. Full context scores 76.1, paragraph selection 74.9, and dense retrieval 71.4. No reported result was found on Lacuna for any pivot or transliteration strategy on LittiChoQA, nor for any cultural-retention measure on it.
All systems share the same stories, the same commercial model and prompt, and the same translation model. Each gets an equal prompt-tuning budget on a held-out development split.
Three ablations isolate the key ingredient, which is choosing the right spans. The first protects randomly chosen spans at the same token budget. The second protects only named entities, found by an off-the-shelf recognizer. The third protects nothing but instructs the translator to keep culture-specific terms, so the translator decides what to keep.
To beat the pivot, the protected pivot must raise $R$ on the cultural questions by a margin outside bootstrap confidence intervals. It must do so while adding no more than a modest fraction of the native-script cost, and without losing accuracy on the matched factual questions. To show the detector matters, the protected pivot must beat random protection and entity-only protection at equal $C$.
Why not just…?
Why not just translate everything to English, since that is cheaper and more accurate? On generic tasks it is, and the diagnostic experiment could simply confirm that. If the plain pivot loses little $R$ on our cultural questions, the constructive claim fails, and we would report that literary cultural content survives translation better than MELLA's results suggest. MELLA's "fluent fool" finding and Kuwanto et al.'s translationese findings are why we expect a real gap. The matched factual-versus-cultural question sets are what decide it.
Why not just prompt the translator to keep culture-specific words? This is the most serious threat. Large translators may already know which words to keep, and if so our detector adds nothing. It is included as an ablation at matched cost. If it matches the protected pivot, the contribution shrinks to the measurement study and the frontier, and we will say so. We expect it to under-protect in low-resource varieties such as Bhojpuri and Awadhi, where the translator may not recognize a term as culture-bearing in the first place.
Why not just fix the tokenizer so native text is cheap? Efficient Low-Resource Language Models Using Tokenizer Transfer removes fragmentation by giving an open model a dedicated tokenizer, and ViT5: Pretrained Text-to-Text Transformer for Vietnamese Language Generation shows language-specific models can beat multilingual ones. These are the right answer for anyone who can retrain a model. They do not apply when the deployer rents a closed commercial model, which is the setting Nag et al. study. As a secondary comparison, we will report a tokenizer-transferred open model's $R$ and $C$ to show where that path dominates.
Why not just retrieve relevant passages instead of sending the whole story? LittiChoQA already reports dense retrieval as its weakest regime, and retrieval shortens context without addressing the language pivot. Our cross of context regime with pivot strategy measures whether the two losses add up, which neither the pivot nor retrieval can reveal alone.
Decisive Checks
The first check is whether the cultural retention gap exists. If the plain English pivot's retention on cultural questions is within noise of native script, the premise is wrong and the protected pivot is moot. The second check is the shape of the frontier. The claim holds only if protecting the top few percent of detected spans closes most of the retention gap, and random protection at the same budget does not. The third check is whether the automatic spans agree with speaker judgement. The detector's spans must overlap substantially with the native-speaker annotations. Speakers must also rate outputs from the protected pivot as more faithful to the story's cultural meaning, not merely as containing more native words. That rating matters because a native word that survives while its meaning is garbled is exactly the kind of surface-level authenticity the heritage literature warns about.
References
- Cost-Performance Optimization for Processing Low-Resource Language Tasks Using Commercial LLMs
- Challenges and Opportunities of Sichuan Opera Face Changing in Haidilao in Cultural Communication
- An Investigation of Yi Folk Song Knowledge and Development in Guizhou, China
- MELLA: Bridging Linguistic Capability and Cultural Groundedness for Low-Resource Language MLLMs
- LittiChoQA: Literary Texts in Indic Languages Chosen for Question Answering
- Mitigating Translationese in Low-resource Languages: The Storyboard Approach
- Efficient Low-Resource Language Models Using Tokenizer Transfer
- ViT5: Pretrained Text-to-Text Transformer for Vietnamese Language Generation