Nhu VoResearch Assistant · NLP & LLMs
Contact
Nhu Vo  /  Datasets

Build the data. Define the task. Evaluate the system.

Three open medical NLP datasets from my PhD research, plus corpora I have led in other projects, from anti-corruption and electric-mobility discourse to learner English. The table compares them at a glance; each card shows my role, the scale, and how quality was checked.

AI Dataset LaboratoryEn-ViMedNERViMedCSSMedEVViACSEV Shift CorpusViLECBusiness Speeches CorpusVnC 1.0
Dataset portfolio

Datasets I built

Scale is the main size measure; evidence is the headline quality check or result.

DatasetTaskScaleEvidenceRoleAccessYear
Open medical NLP datasets
En-ViMedNERBiomedical & cross-lingual NER4,392Abstract pairsκ = 0.83Annotator agreementCo-leadOpen2026
ViMedCSSASR · code-switched speech34Hours of speechκ = 0.65Annotator agreementCo-developerOpen2026
MedEVMachine translation EN↔VI~360KSentence pairs52.21 BLEUBest EN→VI (validation)LeadOpen2024
Other research corpora
ViACSDiscourse analysis · sentiment179,391News articlesκ = 0.744Topic relevance, 98.67% agreedCo-leadRestricted2026
EV Shift CorpusStance · multi-label concerns36,046Commentsκ = 0.804Evaluative orientationCo-leadRestricted2026
ViLECLearner corpus267,512Words—Project leadOn request2024
Business Speeches CorpusBusiness discourse corpus12.5M+Words—Project managerOn request2022
VnC 1.0General corpus · POS, NER100MWords—Project leadOn request2020

Open medical NLP datasets

Built during my PhD; data and papers are public.

A parallel corpus of 4,392 PubMed abstracts for Vietnamese and English→Vietnamese cross-lingual biomedical NER, annotated with 21 UMLS semantic types.

4,392Abstract pairs
44,892Sentence pairs
202,949Aligned entity mentions

Evidence

  • κ = 0.83Annotator agreement
  • 98.32%Entity exact match, LLM label projection
  • 450Human-reviewed test sentences

Construction

  • Source: MedMentions, machine-translated to Vietnamese
  • Translations reviewed by Vietnamese medical experts
  • LLM-assisted label projection, automated consistency checks, manual review of flagged cases

Evaluation

  • Vietnamese biomedical NER and English→Vietnamese cross-lingual NER
  • Compares supervised, multilingual and prompt-based approaches
Notes and related papers
  • Silver-standard corpus; the 450-sentence mini-test is human-reviewed.
  • Paper: En-ViMedNER (Findings of EMNLP 2026)

34 hours of Vietnamese medical speech with English medical terms mixed in: a benchmark for robust ASR on code-switched terms.

34Hours of speech
16,576Utterances
5Medical topics

Evidence

  • κ = 0.65Annotator agreement
  • 500Double-annotated utterances

Construction

  • English terms from a bilingual medical dictionary searched in public Vietnamese medical videos
  • 700+ hours transcribed; about 5% passed the code-switching filters
  • Annotator disagreements led to an automatic boundary-refinement step

Evaluation

  • WER and CS-WER (errors on medical terms only)
  • Hard split: 658 utterances with rare, unseen terms
  • 7 off-the-shelf ASR models, contextual biasing and fine-tuning
Notes and related papers
  • Paper: ViMedCSS (LREC 2026)

About 360K English–Vietnamese medical sentence pairs for medical machine translation.

~360KSentence pairs

Evidence

  • 52.21 BLEUBest EN→VI (validation)
  • 42.66 BLEUBest VI→EN (validation)

Construction

  • Sentences aligned across documents already published in both languages
  • Mismatched pairs filtered out
  • Main sources: Vietnamese medical journal abstracts (over half), then MSD Manuals

Evaluation

  • All four models fine-tuned on MedEV beat Google Translate in both directions
  • Per-genre results: MSD Manuals score highest; terminology-heavy abstracts score lowest
Notes and related papers
  • Paper: Improving Vietnamese-English Medical MT (LREC-COLING 2024)

Other research corpora

Corpora from other research projects. Open a card for method details.

ViACS

Restricted

Vietnamese Anti-Corruption Corpus · Co-lead, 2026

Vietnamese news and social-media texts on anti-corruption, 2017–2025, indexed by province, year and keyword.

179,391News articles
53,798Social-media comments
154News outlets

Evidence

  • κ = 0.744Topic relevance, 98.67% agreed
  • κ = 0.765Geographic tagging, 96.67% agreed
  • 150Double-coded items
Method, facts and papers

Construction

  • 55 keywords from 12 legal and policy documents, LLM-expanded (Gemini 2.5 Pro, GPT-5.1 Pro), kept only if both domain experts accepted
  • Semi-automated crawling of national and provincial outlets, Facebook and YouTube
  • Unicode normalization, noise removal, deduplication; word segmentation for comments

Evaluation

  • Corpus-assisted critical discourse analysis: keyword frequency, lexical density
  • Comment-level sentiment analysis
  • KWIC concordance search in a custom Streamlit app
My role
Co-lead
Year
2026
Published
Under review
Language
Vietnamese
Domain
Governance · media discourse
Modality
Text · news & social media
  • Personal identifiers removed; full news texts restricted to sentence-level concordance lines.
  • Paper: Anti-corruption discourse across news and social media in Vietnam (under review, Critical Discourse Studies)

EV Shift Corpus

Restricted

Vietnamese online discourse on Ho Chi Minh City's electric two-wheeler transition · Co-lead, 2026

Online comments on the 2025 plan to move about 400,000 ride-hailing and delivery motorbikes to electric, from social media, forums and news sites.

36,046Comments
1,879Posts and news articles
69News sites

Evidence

  • κ = 0.804Evaluative orientation
  • F1 = 0.934Concern labels (weighted)
  • 600Validation comments
Method, facts and papers

Construction

  • Crawled Facebook, YouTube, TikTok, X, Threads, five forums (VOZ, Otofun, Otosaigon, Tinhte, VinFast Vietnam) and 69 news sites; policy-relevance screening
  • Rule-based and LLM-assisted normalization of slang, abbreviations and teencode
  • Deduplication, sentence segmentation (SAT-3L), word segmentation (VnCoreNLP)

Evaluation

  • Labels: policy-directed evaluative orientation and five concerns (cost, infrastructure, safety, environment, policy trust)
  • Human-validated LLM labelling: GPT-5.2 vs raters κ 0.792 / 0.913, F1 0.948 / 0.974
  • Baselines: VnEmoLex lexicon, PhoBERT sentiment
My role
Co-lead
Year
2026
Published
Under review
Language
Vietnamese
Domain
Urban mobility · policy discourse
Modality
Text · social media, forums, news comments
  • Preprocessing checked by two researchers on 9,016 sentences.
  • Usernames, account IDs and profile data removed.
  • Paper: Data-Driven Equity-Sensitive Urban Mobility Planning (under review, Journal of Urban Mobility)

ViLEC

On request

Vietnamese Learner English Corpus · Project lead, 2024

English written by Vietnamese students across five courses.

267,512Words
15,129Sentences
258Documents
Method, facts and papers

Construction

  • Student writing collected at the Faculty of Foreign Languages, Nguyen Tat Thanh University (2021–2022)

Evaluation

  • Corpus-based study of learner English and pedagogical applications
My role
Project lead
Year
2024
Published
Under review
Language
English (learner) · Vietnamese L1
Domain
Education
Modality
Text
  • Paper: ViLEC (under review, SAGE Open)

Business Speeches Corpus

On request

Corpus of transcribed business talks · Project manager, 2022

Transcribed business talks, built at VinUniversity with Dr Thuy Bui and Prof. Rohit Verma.

12.5M+Words
720K+Sentences
1,622Talks
Method, facts and papers

Construction

  • Built at VinUniversity (2022–2023)

Evaluation

  • Corpus-based research on business discourse
My role
Project manager
Year
2022
Language
English
Domain
Business
Modality
Speech transcripts

VnC 1.0

On request

Vietnamese Corpus 1.0 · Project lead, 2020

A general Vietnamese corpus (10% spoken, 90% written), word-segmented, POS- and NER-tagged.

100MWords
Method, facts and papers

Construction

  • Built at the Computational Linguistics Center, VNUHCM – University of Science (2020–2021)
  • Sampling criteria published at FAIR 2020

Evaluation

  • General-purpose resource for Vietnamese corpus linguistics, lexicography and NLP
My role
Project lead
Year
2020
Published
FAIR 2020
Language
Vietnamese
Domain
General · spoken & written
Modality
Text
  • Paper: Language sampling criteria for Vietnamese corpus construction (FAIR 2020)
Dataset pipeline

From source material to model evaluation

The steps I follow when building a dataset. Each project adapts them.

  1. 01Raw data

    Source corpora, speech, and parallel text

  2. 02Cleaning

    Preparation, alignment, normalization

  3. 03Annotation

    Label definitions and projection

  4. 04Quality control

    Consistency checks and expert review

  5. 05Benchmark

    Task definition, splits, baselines

  6. 06AI application

    Evaluated models and prototypes