Text Generation in Low-Resource Languages Using Human-Curated Vocabularies

• Human-curated vocabularies cover more languages than common text resources and record words, meanings, and other linguistic information.Source: §1 · p. 1
• LLM-based synthetic generation relies on knowledge encoded during pretraining, limiting its effectiveness for languages poorly represented in that data.Source: §1 · p. 1
• The paper asks whether LLMs can use human-curated vocabularies to generate coherent, lexically diverse text in low-resource and extremely underrepresented languages.Source: §1 · p. 2
• The training-free framework combines LLM reasoning, human-curated vocabularies, few-shot human-written exemplars, and Zipfian sampling.Source: §1 · p. 2
• Across six languages, the framework improves lexical diversity and reduces off-target generation over practical baselines, while human ratings show fluency gains, especially for Acehnese and Hitu.Source: §1 · p. 2

• LLM-generated synthetic text can expand pretraining and post-training resources, but tends to have lower lexical diversity than human-written data.Source: §2 · p. 2
• Low-resource generation is difficult because linguistic knowledge is limited, quality can deteriorate as resources decrease, and models may drift into higher-resource languages.Source: §2 · p. 2
• Prior work uses human-curated lexical resources for augmentation, word-for-word synthesis, translation, and contextual support for LLM translation.Source: §2 · p. 2
• Unlike related work that translates or modifies text, this paper generates new text directly in the target language, using vocabularies as lexical knowledge and exemplars as syntactic guidance.Source: §2 · pp. 2–3

• The pipeline collects dictionaries and frequency information, gathers sentence and paragraph exemplars, samples candidate words, and assembles a generation prompt.Source: §3 · p. 3
• Dictionary entries include a target-language word, English equivalent and definition, part-of-speech tag, and sometimes an example sentence.Source: §3 · p. 3
• Because target-language frequency estimates are unavailable or unreliable, dictionary words inherit estimated frequencies from their English equivalents; entries without frequencies are excluded.Source: §3 · p. 3
• Native-speaker annotators write 40 sentences and 5 paragraphs per language; these examples provide syntactic and contextual guidance and link text to dictionary entries.Source: §3 · p. 3
• Candidates are sampled by estimated frequency rather than uniformly, avoiding overrepresentation of rare words and underrepresentation of common words, including function words.Source: §3 · p. 3
• The method divides ranked vocabulary into high- and low-frequency sets, samples low-frequency words from a topic-related subset, and combines both sets into k candidates.The configuration uses k = 300 candidates, s = 3 exemplars, and topic fraction ρ = 0.25.Source: §3 · pp. 3–4
• The prompt pairs each exemplar with its associated dictionary entries, then adds a newly sampled candidate set and instructs the reasoning-enabled LLM to generate text using suitable words.Source: §3 · p. 4

4.1 Configurations

• The study evaluates six Southeast Asian languages spanning medium, low, and extremely low resource availability, using three LLMs, two thinking modes, three prompts, and five text lengths.The 540 configurations each produce 100 outputs, for 54,000 total.Source: §4.1 · pp. 4–5
• The three prompts are topic-conditioned zero-shot, few-shot exemplars only, and the proposed prompt combining exemplars with dictionary word candidates.Source: §4.1 · p. 5

4.2 Evaluation Metrics.

• Human fluency ratings cover three models with thinking enabled, three prompts, and three text lengths, using 30 texts per configuration and two annotators per language.Annotators rate naturalness on a 1–5 scale; each text’s score is the mean of the two ratings.Source: §4.2 · p. 5
• Off-target rate is the proportion of generated words absent from the target-language dictionary, with higher values indicating greater language confusion.Source: §4.2 · p. 5
• Lexical diversity measures distinct valid target-language n-grams relative to all n-grams, averaged over lengths 1 through 5.The score decreases with repeated n-grams and off-target words.Source: §4.2 · p. 5
• Table 2 reports fluency, off-target rate, and lexical diversity by language and prompt, averaging across models and thinking modes, except fluency uses thinking-enabled models.Source: Table 2 · p. 6

Fluency

• Our prompt leads human-rated fluency in Minangkabau, Acehnese, and Hitu by 0.33, 0.69, and 0.79 points over the best baselines, respectively.For Indonesian, prompts score near the maximum; for Vietnamese and Javanese, our prompt trails both baselines, particularly on Vietnamese paragraphs (3.54 vs. 4.72 for zero-shot). The authors hypothesize dictionary quality may contribute.Source: §5 · p. 6

Off-Target Rate

• Off-target rates stay at or below 0.090 across languages; Hitu falls from 0.836 zero-shot to 0.060, and Acehnese from 0.315 few-shot to 0.064.Source: §5 · p. 6

Lexical Diversity

• Lexical diversity is highest for every language and text length; Indonesian scores 0.72–0.87 versus baselines’ 0.43–0.61.For Hitu, our prompt scores 0.51–0.75 versus baselines below 0.09 at every length, averaging 0.57 higher than few-shot prompting; diversity declines with text length.Source: §5 · pp. 6–7

Dictionary Qualities Hinder Fluency in Vietnamese and Javanese

• Dictionary quality constrained fluency in Vietnamese and Javanese because uncommon or dialect-mismatched valid words sounded unnatural to annotators.The authors recommend checking both everyday lexical coverage and alignment with the intended language variety before large-scale generation.Source: §6 · p. 7

Reliability of the LLM-as-a-Judge for Fluency Evaluation

• The fluency judge’s reliability varies by language: Hitu has the weakest Spearman correlation (0.15) and highest MAE (+1.19), while scores tend to overestimate human ratings in several languages.Overestimation is especially pronounced for Minangkabau (+0.60) and Acehnese (+0.58); the authors caution against treating the judge as a substitute for human evaluation.Source: §6 · pp. 7–8

Thinking vs. Non-thinking Mode

• Thinking mode usually improves generation quality, but its effect is model-dependent: Gemini 3.5 Flash gains most, while Claude Haiku 4.5 shows no benefit.For Gemini 3.5 Flash, off-target rate falls by 0.06 and lexical diversity rises by 0.03; Table 4 averages results across three languages and prompts.Source: Table 4 · p. 7, §6 · p. 8

Robustness Under Increasing Generation Volume

• Across prompting strategies, the percentage of unique valid n-grams decreases as more sentences are generated, while the proposed prompt remains higher than baselines in both Minangkabau and Hitu.At 100 versus 1,000 sentences, Minangkabau zero-shot/few-shot scores change from 0.32/0.36 to 0.14–0.15; Hitu scores change from 0.16/0.05 to 0.09/0.04.Source: §6 · p. 8

7 Conclusion

• The training-free framework combines human-curated vocabularies, human-written exemplars, and LLM reasoning to generate synthetic text in low-resource languages.The framework samples candidate words according to frequency and constructs prompts from sampled words and exemplars.Source: §7 · p. 8
• Across six languages, the framework produces more lexically diverse text with fewer off-target words than baselines, while improving fluency particularly for Acehnese and Hitu.The authors describe these results as showing the potential of human-curated vocabularies for bootstrapping synthetic corpora where text resources are limited.Source: §7 · p. 8
• Dictionary quality and alignment with the target language variety can substantially affect generation quality, so vocabulary validation matters before large-scale generation.Providing word candidates and exemplars also requires additional input tokens, creating a trade-off between generation quality and cost.Source: §7 · p. 8
• Future work should extend the framework to more low-resource languages and investigate whether its synthetic corpora improve LLM pretraining and post-training.The study evaluates generated-text fluency, off-target rate, and lexical diversity, but does not train or fine-tune models on the generated data.Source: §7 · pp. 8–9

Limitations

• The experiments cover six Southeast Asian languages, all written in Latin script, and cannot establish generalization to non-Latin scripts or agglutinative or polysynthetic morphology.Five languages are Austronesian, and five are spoken in Indonesia and share considerable lexical material with Indonesian, a setting the authors say is likely to favor positive transfer.Source: Limitations · p. 9
• Hitu’s status as unseen during pretraining is inferred from model behavior, not established, because the commercial models’ training corpora are not public.Source: Limitations · p. 9
• The study uses three proprietary economy-tier models accessed through commercial APIs, so it remains unclear whether the gains hold for larger or more recent models.Generating 54,000 texts made broader model coverage too costly; the findings are restricted to the latest versions of those model families at the time of the experiments.Source: Limitations · p. 9
• The evaluation measures fluency, off-target rate, and lexical diversity, but does not test whether training or fine-tuning on the generated data benefits downstream models.The authors identify downstream evaluation, particularly for low-resource languages, as future work.Source: Limitations · p. 9

Ethical Considerations

• Native speakers of the six target languages wrote exemplars and rated generated texts; they were informed about the study, paid above regional minimum wage in Indonesia and Vietnam, and could withdraw.No personal or sensitive information was collected or included in released materials.Source: Ethical Considerations · p. 9

A Dictionary Resources

• Dictionary preparation varied by resource: Javanese and Minangkabau entries were scraped, Hitu and Acehnese PDFs were extracted, and Indonesian and Vietnamese datasets were processed into a consistent format.Non-English definitions were translated into English; some missing Javanese fields were filled using Gemini 3.5 Flash.Source: A Dictionary Resources · p. 11

B Few-Shot Exemplar Creation Procedure

• One native or fluent local-language annotator per language had to have lived in the target-language region; each created 40 sentences and five paragraphs.The instructions included ten sentences of approximately five words.Source: B Few-Shot Exemplar Creation Procedure · pp. 11–12, Figure 4 · p. 12

C Hyperparameter Settings

• Hyperparameters were selected through preliminary experiments in which annotators rated outputs from configurations generating 15 five-word sentences and paragraphs.Table 6 lists the settings and values searched.Source: C Hyperparameter Settings · p. 12, Table 6 · p. 12

D Topic-Conditioned Zero-Shot Prompt

• The topic-conditioned zero-shot prompt specifies the target language, output type, topic, and sentence length, using one of 20 listed topics.Source: D Topic-Conditioned Zero-Shot Prompt · p. 12, Figure 5 · p. 12

E Few-Shot Exemplars Only Prompt

• In the exemplars-only prompt, examples guide style, grammar, fluency, and natural sentence structure, but outputs must not reuse their main semantic content.The prompt requests one coherent target-language sentence or paragraph.Source: E Few-Shot Exemplars Only Prompt · p. 13, Figure 6 · p. 13

F Few-Shot Exemplars and Word Candidates Prompt

• The word-candidate prompt presents entries with the target-language word, part of speech, meaning, and example usage, and treats candidates as guidance rather than a strict requirement.It instructs the model to use candidates where they fit naturally and prioritize coherent, meaningful output.Source: F Few-Shot Exemplars and Word… · p. 14

G Human Evaluation Guidelines

• Native speakers rated each text’s fluency on a 1–5 Likert scale, with rubric descriptions ranging from unintelligible or error-filled text to fluent, natural text.Source: G Human Evaluation Guidelines · pp. 14–15, Figure 8 · p. 15

H Output Sample Comparison Between Baseline and Our Pipeline

• The passage attributes the zero-shot model’s failure on Hitu sentences and paragraphs to its limited knowledge of the language.The same model generates more fluent Hitu text when given exemplars and vocabulary options, without further training.Source: H Output Sample Comparison Between… · p. 15

I Information About AI Usage

• The authors used AI assistants for writing polish and literature search, manually verified papers found through AI-assisted search, and state that all ideas originate from the authors.Source: I Information About AI Usage · p. 15

H Output Sample Comparison Between Baseline and Our Pipeline

• With Claude Haiku 4.5, few-shot exemplars and word candidates produce more fluent Hitu text than zero-shot, which fails to generate Hitu sentences and paragraphs.Table 7 presents sample outputs for the baseline and method at the five-word sentence and paragraph levels.Source: Table 7 · p. 15, H Output Sample Comparison Between… · p. 15

J Detailed Results by Model and Thinking Mode

• Table 8 reports fluency, off-target rate, and lexical diversity for each language, model, thinking mode, and prompt.The supplied passages do not include the table's cell values.Source: J Detailed Results by Model and… · p. 15, Table 8 · p. 18
• The table covers both medium-resource and extremely low-resource languages.Its entries continue across pages for each resource group.Source: J Detailed Results by Model and… · pp. 16–18
• Fluency was human-evaluated only for models with thinking enabled, and bold marks the best result per model and thinking mode.Source: Table 8 · p. 18