ViDia2Std: A Parallel Corpus and Methods for Low-Resource Vietnamese Dialect-to-Standard Translation

Khoa Anh Ta; Nguyen Van Dinh; Kiet Van Nguyen

cs.CLMar 10, 2026

ViDia2Std: A Parallel Corpus and Methods for Low-Resource Vietnamese Dialect-to-Standard Translation

Authors: Khoa Anh Ta, Nguyen Van Dinh, Kiet Van Nguyen

Organizations: Faculty of Information Science and Engineering, University of Information Technology, Ho Chi Minh City, Vietnam · Vietnam National University, Ho Chi Minh City, Vietnam

Abstract

Vietnamese exhibits extensive dialectal variation, posing challenges for NLP systems trained predominantly on standard Vietnamese. Such systems often underperform on dialectal inputs, especially from underrepresented Central and Southern regions. Previous work on dialect normalization has focused narrowly on Central-to-Northern dialect transfer using synthetic data and limited dialectal diversity. These efforts exclude Southern varieties and intra-regional variants within the North. We introduce ViDia2Std, the first manually annotated parallel corpus for dialect-to-standard Vietnamese translation covering all 63 provinces. Unlike prior datasets, ViDia2Std includes diverse dialects from Central, Southern, and non-standard Northern regions often absent from existing resources, making it the most dialectally inclusive corpus to date. The dataset consists of over 13,000 sentence pairs sourced from real-world Facebook comments and annotated by native speakers across all three dialect regions. To assess annotation consistency, we define a semantic mapping agreement metric that accounts for synonymous standard mappings across annotators. Based on this criterion, we report agreement rates of 86% (North), 82% (Central), and 85% (South). We benchmark several sequence-to-sequence models on ViDia2Std. mBART-large-50 achieves the best results (BLEU 0.8166, ROUGE-L 0.9384, METEOR 0.8925), while ViT5-base offers competitive performance with fewer parameters. ViDia2Std demonstrates that dialect normalization substantially improves downstream tasks, highlighting the need for dialect-aware resources in building robust Vietnamese NLP systems.

Figures & tables

Table 1: Comparison of Translation Models on Dialectal and Normalized Data.

Figure 1: Overview of our dataset creation pipeline, from raw social media data to the final dialect-standard parallel corpus.

Table 2: Sample sentence pairs: dialect vs. standard

Region	Number of Sentences	Avg. Words
		Dialect	Standard
Central	9,033	10.55	10.82
Northern	1,054	9.31	9.28
Southern	3,570	10.58	11.02
Total	13,657	10.46	10.77

Table 3: Basic Dataset Statistics

Model	ROUGE-L	BLEU	METEOR	WER	CER	Parameters
BARTpho-word-base	0.9167	0.7601	0.8625	0.1527	0.1049	150M
ViT5-base	0.9300	0.7934	0.8802	0.1340	0.0876	310M
BARTpho-syllable-base	0.9218	0.7597	0.8627	0.1513	0.1045	132M
Vietnamese-correction-v2	0.9257	0.7723	0.8746	0.1416	0.0988	396M
mBART-large-50	0.9384	0.8166	0.8925	0.1226	0.0754	611M

Table 4: Intrinsic Evaluation Results

Method	Total Samples	Accepted	Unaccepted	Acceptance Rate (%)
Microsoft Azure AI Translator	600	68	532	11.33
Microsoft Azure AI Translator + Normalized	600	134	466	22.33
Google Cloud Translation – Basic	600	230	370	38.33
Google Cloud Translation – Basic + Normalized	600	275	325	45.83
DeepSeek-V3	600	327	273	54.50
DeepSeek-V3 + Normalized	600	379	221	63.17

Table 5: Machine Translation Evaluation Results (Overall)

Threshold ( $\geq$ )	Improved	Worsened	Ratio
1	316	186	1.70
2	152	48	3.17
3	69	17	4.06
4	32	1	32.00
5	11	0	$\infty$
6	3	0	$\infty$

Table 6: Consistency of Normalization Impact Across 6 Systems

System	Improved	Worsened
Microsoft Azure AI Translator	88	22
Kimi-K2-Instruct	125	48
Gemini 1.5 Flash	102	42
Google Cloud Translation	88	43
DeepSeek-V3	102	50
Gemini 2.0 Flash	78	47

Table 7: Per-System Sentence Improvement vs. Worsening

Error Category	Count (%)	Description
MT Model Fragility	118 (46.8%)	The MT system fails to translate a correctly normalized sentence.
LLM Evaluator Noise	90 (35.7%)	The evaluator provides inconsistent ratings for translations of equivalent quality.
Normalization Model Error	44 (17.5%)	The normalization model incorrectly alters the meaning of the source sentence.
Total	252 (100%)

Table 8: Categorization of Translation Regression Causes Across All Systems (percentages shown in parentheses next to counts)

	Before			After
Class	P	R	F1	P	R	F1
NEGATIVE	0.89	0.45	0.59	0.86↓	0.62↑	0.72↑
NEUTRAL	0.47	0.30	0.37	0.48↑	0.42↑	0.45↑
POSITIVE	0.32	0.87	0.47	0.44↑	0.82↓	0.58↑
Accuracy	0.51			0.62↑
Macro Avg	0.56	0.54	0.48	0.59↑	0.62↑	0.58↑

Table 9: Sentiment analysis before vs. after dialect normalization. Bold↑ indicates improvement; ↓ indicates decline.

Type of Change	Count	Percentage (%)
Incorrect $\to$ Correct	265	16.53%
Correct $\to$ Incorrect	80	4.99%
Correct $\to$ Correct	731	45.60%
Incorrect $\to$ Incorrect	527	32.88%
Total	1603	100.00%

Table 10: Comparison: Improvement vs. Regression after Normalization

Low-Resource MT

Low-resource machine translation develops translation systems for languages with scarce parallel text, tools, or standardized evaluation.

See all 87 papers

How Robust Are LLMs to Vietnamese Dialects?
Aug 11, 2026Minh Tran, Trinh Chau, Thanh-Nhan Le +4LLM Evaluation Language Model Robustness
University of Science, VNU-HCM · VNU University of Engineering and Technology · New York University Abu Dhabi +4
Phonetic Modeling of Dialectal Variation in Vietnamese Speech
May 23, 2026Quan Ngoc Hoang, Long Hoang Huu Nguyen, Nghia Hieu Nguyen +2Speech Processing Automatic Speech Recognition
4Vietnam National University, Ho Chi Minh city, Vietnam
Unified Multi-Dialectal Neural Machine Translation for Bangla Using the Dwadash Benchmark Corpus
Aug 12, 2026Rakib Ullah, Md. Ruhul Islam, Tanbir Ahmed +1Low-Resource MT Low-Resource Language Processing
Department of Computer Science and Engineering SYLHET ENGINEERING COLLEGE, SYLHET Sylhet, Bangladesh

Aug 11, 2026J/K move · Enter open · S save

How Robust Are LLMs to Vietnamese Dialects?

Minh Tran, Trinh Chau, Thanh-Nhan Le, Nam Tran+3

University of Science, VNU-HCM · VNU University of Engineering and Technology · New York University Abu Dhabi+4

LLM Evaluation Language Model Robustness

Open paper AskPDF arXiv

Large Language Models (LLMs) are typically evaluated on standard written Vietnamese, yet everyday communication frequently involves regional dialects that preserve meaning but differ in surface form. Existing Vietnamese dialect work largely addresses this issue through dialect-to-standard normalization instead of measuring how the model fails under Vietnamese dialectal inputs. To address this gap, we present the first systematic evaluation of LLM robustness to Vietnamese dialect variation across multiple tasks, quantifying performance degradation and failure patterns. We introduce VialectBench (Vietnamese Dialects Benchmarking), a controlled benchmark for testing whether model decisions remain stable across six Vietnamese dialect groups. VialectBench contains 400 Standard Vietnamese source instances and 2,400 human-written dialectal rewrites spanning emotion recognition (ER), natural language inference (NLI), question answering (QA), and multiple-choice question answering (MCQA). Dataset evaluation with a fixed reference language model shows that the dialectal rewrites induce a measurable model-relative likelihood shift while remaining nearly equal in length to their Standard counterparts. Across ten instruction-tuned models, dialectal inputs reduce average performance by 2.82%, and no evaluated model is fully dialect-invariant. All four tasks are affected, with QA showing the largest average degradation. Robustness also varies substantially across dialect groups: PNT3 and PNT2 cause the largest average performance drops, at 6.17% and 4.73%, respectively, whereas PNB slightly improves average performance by 0.42%. The Central dialect group (PNT1-PNT4) also yields the highest average harmful-flip rate across all models, at 6.54%. These findings show that strong performance on Standard Vietnamese does not guarantee reliable behavior under meaning-preserving regional variation.

Abstract

Low-Resource MT

Explore similar work

How Robust Are LLMs to Vietnamese Dialects?

Phonetic Modeling of Dialectal Variation in Vietnamese Speech

Unified Multi-Dialectal Neural Machine Translation for Bangla Using the Dwadash Benchmark Corpus