Unfragmenting Nawatl NMT: Single-Variant Specialization for Masewalkopa
Organizations: Department of Linguistics, University of Colorado Boulder
Abstract
In minoritized (low-resource) language technologies, a common response to data sparsity is to aggregate text across multiple varieties and treat the combined data as a single corpus. However, mixed corpora of closely related indigenous variants can produce an artificial register that is accountable to no single speaker community, a structural failure we call variant fragmentation. This issue is acute for minoritized languages native to Mexico, such as Nawatl. Building on a 70,000-pair multi-variant foundation, we specialize an mT5-large checkpoint to Masewalkopa (nhn), the Nawatl variety spoken in Santa Ana Tlacotenco. We curate 2,665 parallel pairs, reserve all sentence-level data for evaluation, and incorporate 192 elder-transcribed target sentences via asymmetric backtranslation. A custom finite-state analyzer reduces unanalyzed surface forms from 82.0% to 19.9% and supplies functional syntax tags as plain-text prefixes, allowing the model to reuse existing subword embeddings rather than relying on frozen special tokens. The resulting LoRA-v9c-b model achieves 58.46 dev chrF++ and 58.20/57.90 directional test chrF++. Finally, CTranslate2 int8 quantization reduces CPU latency from 137.5 to 1.8-2.6 seconds per request, enabling cost-bounded Cloud Run deployment.
Figures & tables
| Subset | Pairs | Share | Illustrative Examples | Primary Experimental Role |
|---|---|---|---|---|
| Lexical / short ( tokens) | 1,188 | 44.6% | sesek frío ; nimitstlasohkamate te agradezco | Training coverage, citation forms, short-form deployment probes |
| Sentence level | 1,477 | 55.4% | tlamach tikayawaskeh juntos en comunidad venceremos | Training plus exclusive source of development and test items |
| Total unique | 2,665 | 100.0% | Masewalkopa specialization corpus | |
| Class | Function in the analyzer | Representative constraint |
|---|---|---|
| A-Stem | Stems whose citation or inflected path is organized around an a -final alternation | Restricts which aspect and derivational arcs can follow |
| I-Stem | i -final verbal or predicate paths | Preserves class-specific boundary alternations |
| OWA-Stem | Productive verbal stems associated with an owa pattern | Connects to verbal person and aspect networks |
| ITL-Stem | Nominal stems selecting an -itl -type absolutive realization | Blocks incompatible absolutive suffixes |
| TLE-Stem | Nominal stems selecting a -tle -type path in local orthography | Separates local surface behavior from generic noun rules |
| LI-Stem | Nominal stems selecting -li | Supports possessed and predicate-nominal alternations |
| Development stage | Vocabulary | No analysis | Passthrough | Genuine segmentation |
|---|---|---|---|---|
| Initial baseline | 2,736 | 82.0% (2,244) | 7.9% (216) | 10.1% (276) |
| Post aspect/POS audit | 2,807 | 26.4% (742) | 7.4% (207) | 66.2% (1,858) |
| Batch 22–23 expansion | 2,807 | 25.2% (707) | 7.2% (202) | 67.6% (1,898) |
| Batch 24–26 closure | 2,807 | 19.9% (559) | 7.2% (202) | 72.9% (2,046) |
| Surface | FST decomposition | IGT gloss | Functional projection |
|---|---|---|---|
| tla | tla | OPT/conditional particle | OPT or ADV |
| nichnamikis | ni-ch-namiki-s | 1SG.SBJ-OBJ-meet-FUT | SUBJ OBJ HEAD |
| tokonetsin | to-kone-tsin | 1PL.POSS-child-HON | NP.POSS |
| tisowatl | ti-sowatl | 2SG.SBJ-woman.PRED | SUBJ COP HEAD |
| kwalle | kwalle | good/well | ADV or predicate HEAD |
| ammo | ammo | NEG | NEG |
| Run | Data or Intervention | Epochs | Dev chrF++ | Test es nah | Test nah es |
|---|---|---|---|---|---|
| Baseline Checkpoints | |||||
| Base | Mixed-variant base checkpoint | – | – | 23.91 | 29.54 |
| Baseline LoRA reference | Initial scheduled LoRA snapshot | – | 43.87 | – | – |
| Schedule and Data Augmentation | |||||
| lora_v1 | Real pairs only | 3 | 40.26 | 41.76 | 43.46 |
| lora_v2 | Real pairs, extended schedule | 8 (best 6) | 46.18 | 45.88 | 45.58 |
| Spanish source | Mixed-variant base | Specialized base ( lora_v2 ) | +BT +Dense syntax ( lora_v9c_b ) | Human reference |
|---|---|---|---|---|
| Te agradezco. | tlazohcamati | nimitstlaskamati | nimitstlasohkamate | nimitstlasohkamate |
| Nuestro hijito está bien. | to piltzin cualli | tokone kwalle | tokonetsin kwalle | tokonetsin kwalle |
| Eres una mujer buena. | ti cihuatl cualli | tisowatl kwalle | tisowatl kwalle | tisowatl kwalle |
| Encontraré a nuestro hijito. | ni…namiki topiltzin | niknamikis tokone | nichnaamikis tokonetsin | nichnaamikis tokonetsin |
| Spanish Source | Production Output ( lora_v9c_b ) | Human Reference | Error Diagnosis |
|---|---|---|---|
| sí sé, pero no me acuerdo | kemma nikmate, wan ammo nikmati | kemma nikmate, wan ammo nikilnaamike | Semantic repetition: Loops base cognition verb ( mate ) rather than retrieving psychological memory root ( ilnaamiki ). |
| ayer te vi en la calle | yalwa onimitsitak ipan kalohtle | yalhwa onimitsihtak ipan kalohtle | Orthographic bleed: Omits local Santa Ana aspiration marker ( -h- ) in both time adverbial ( yalhwa ) and root ( ihta ). |
| mañana yo te enseñaré | mustla nimitsmachtia | mustla nimitsmachtis | Aspect truncation: Drops future suffix ( -s ) under complex prefixation, reverting to default non-past aspect ( -machtia ). |
| Runtime | Weight precision | Beam | Latency per sentence |
|---|---|---|---|
| PyTorch CPU | fp32 | 4 | 137.5 s |
| CTranslate2 CPU | int8 | 4 | 1.8–2.6 s, but short-form failures |
| CTranslate2 CPU | int8 | 8 | 1.8–2.6 s, validated production setting |
Machine Translation
Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.
See all 254 papers