🇻🇳 Vietnamese NLP Tasks — Benchmark & SOTA Overview

📈 This page tracks major Vietnamese NLP datasets and models for Dependency Parsing, Intent Detection, Machine Translation, NER, POS Tagging, Semantic Parsing, and Word Segmentation.

Dependency Parsing

🗂️ VnDT v1.1/v1.0: Benchmark treebank >10K sentences.
Test: 1,020 (v1.1), Dev: 200, Rest: Train.

VnDT v1.1

Model	LAS	UAS	Paper	Code
PhoNLP (2021)	79.11	85.47	PhoNLP	Official
PhoBERT-base (2020)	78.77	85.22	PhoBERT	Official
Biaffine (2017)	74.99	81.19	Biaffine Parsing
VnCoreNLP (2018)	71.38	77.35	VnCoreNLP	Official

VnDT v1.0 (Gold POS)

Model	LAS	UAS	Paper	Code
VnCoreNLP (2018)	73.39	79.02	VnCoreNLP	Official
BIST BiLSTM graph (2016)	73.17	79.39	BIST Parser	Official
MSTparser (2006)	70.29	76.47	MSTparser

Intent Detection & Slot Filling

🛫 PhoATIS Dataset (flight booking domain): Train: 4,478, Dev: 500, Test: 893

Model	Intent Acc.	Slot F1	Sent. Acc.	Paper	Code
JointIDSF (2021)	97.62	94.98	86.25	JointIDSF	Official
JointBERT+PhoBERT	97.40	94.75	85.55	JointIDSF	Official

Machine Translation

🌐 PhoMT Dataset: 3.02M sentence pairs | 6 domains (TED, WikiHow, MediaWiki, OpenSubtitles, News, Blog)

Model	EN→VI (BLEU)	VI→EN (BLEU)	Paper	Code
mBART (2020)	43.46	39.78	mBART	Link
Transformer-big	42.94	37.83	Transformer	Link

📋 IWSLT2015: 150K sentence pairs (EN↔VI) | Data & Scripts

Model	BLEU	Paper	Code
Nguyen & Salazar (2019)	32.8	Transformers w/o Tears	Official
Provilkov et al. (2019)	33.27 (uncased)	BPE-Dropout
Xu et al. (2019)	31.4	Layer Norm	Official
Transformer (2017)	28.9	Transformer	Link

Named Entity Recognition (NER)

🩺 PhoNER_COVID19: 10 types, 34,984 entities, 10,027 sentences

Model	F1	Paper	Code
PhoBERT-large	94.5	PhoBERT	Official
XLM-R-large	93.8	XLM-R	Official
BiLSTM-CRF + CNN-char	91.0	BiLSTM-CRF	Link

📄 VLSP 2016 NER: 16,861 train/dev, 2,831 test sentences.

Model	F1	Paper	Code
PhoBERT-large	94.7	PhoBERT	Official
PhoNLP	94.41	PhoNLP	Official
vELECTRA	94.07	vELECTRA	Official
VnCoreNLP	91.30	VnCoreNLP	Official

Part-of-Speech Tagging

🔤 VLSP 2013: 27,870 train/dev, 2,120 test

Model	Accuracy	Paper	Code
PhoBERT-large	96.8	PhoBERT	Official
vELECTRA	96.77	vELECTRA	Official
PhoNLP	96.76	PhoNLP	Official
PhoBERT-base	96.7	PhoBERT	Official
VnCoreNLP-VnMarMoT	95.88	VnMarMoT	Official
BiLSTM-CRF + CNN-char	95.40	BiLSTM-CRF	Official
RDRPOSTagger	95.11	RDRPOSTagger	Official

Semantic Parsing

🗃️ ViText2SQL: 10K question/SQL pairs, the first public Text-to-SQL dataset for Vietnamese.

Model	Exact Match Acc.	Paper	Code	Note
IRNet (2019)	53.2	ViText2SQL	Link	Using PhoBERT encoder
EditSQL (2019)	52.6	ViText2SQL	Link	Using PhoBERT encoder

Word Segmentation

✂️ VLSP 2013: 75k train, 2,120 test sentences (manually word-segmented)

Model	F1	Paper	Code
UITws-v1 (2019)	98.06	UITws-v1	Official
VnCoreNLP-RDRsegmenter (2018)	97.90	VnCoreNLP	Official
UETsegmenter (2016)	97.87	UETsegmenter	Official
vnTokenizer (2008)	97.33	vnTokenizer
JVnSegmenter (2006)	97.06	JVnSegmenter