How Robust Are LLMs to Vietnamese Dialects?
TL;DR Evaluates how reliably large language models handle regional Vietnamese dialect variation, revealing robustness gaps that matter for inclusive Vietnamese NLP.
A Multi-Task Approach Towards Robust Vietnamese Audio-Based Toxic Span Detection
TL;DR A multi-task approach jointly detects toxic Vietnamese speech and localizes its harmful spans, improving robustness through shared supervision.
Don’t Take it Literally! Idiom-aware Vietnamese Translation via In-context Learning
TL;DR An in-context learning framework helps LLMs translate Vietnamese idioms by accounting for figurative and culturally grounded meaning.
ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances
TL;DR ViToSA introduces the task of locating toxic spans directly in Vietnamese speech, together with data and baseline models.
ZeST: A Zero-resourced Speech-to-Speech Translation Approach for Unknown, Unpaired, and Untranscribed Languages
TL;DR ZeST translates speech between languages without transcripts or parallel data by using visually grounded self-supervised speech representations.
Multi-Dialect Vietnamese: Task, Dataset, Baseline Models and Challenges
TL;DR A new dataset and evaluation setting expose how current Vietnamese NLP models struggle across regional dialects.
ViHateT5: Enhancing Hate Speech Detection in Vietnamese with Unified Text-to-Text Transformer Models
TL;DR ViHateT5 reframes Vietnamese hate-speech tasks as text-to-text generation, providing unified models and resources for language-safety research.
VGSAlign: Bilingual Speech Alignment of Unpaired and Untranscribed Languages using Self-Supervised Visually Grounded Speech Models
TL;DR VGSAlign discovers bilingual speech correspondences without transcripts by aligning visually grounded representations across languages.
SMTCE: A Social Media Text Classification Evaluation Benchmark and BERTology Models for Vietnamese
TL;DR SMTCE consolidates Vietnamese social-media classification tasks into a common benchmark and evaluates BERT-based models across them.
Constructive and Toxic Speech Detection for Open-Domain Social Media Comments in Vietnamese
TL;DR The ViCTSD dataset supports joint study of constructive and toxic language in open-domain Vietnamese social-media comments.
† Undergraduate student supervised by Nguyen Thanh Luan.