Publications

2025

1. Pre-print

   CAMELLIA : BENCHMARKING CULTURAL BIASES IN LLMS FOR ASIAN LANGUAGES

   Tarek Naous(1) , Anagha Savit(1) , Carlos Rafael Catalan(2) , Geyang Guo(1) , Jaehyeok Lee(3) , Kyungdon Lee(3) , Lheane Marie Dizon(2) , Mengyu Ye(4) , Neel Kothari(1) , Sahajpreet Singh(5) , Sarah Masud(6) , Tanish Patwa(1) , Trung Tanh Tran(7) , Zohaib Khan(8) , Alan Ritter(1) , JinYeong Bak(3) , Keisuke Sakaguchi(4) , Tanmoy Chakraborty(9) , Yuki Arase(10), Wei Xu(1)

   (1) Georgia Institute of Technology, (2) Samsung R&D Institute Philippines, (3) Sungkyunkwan University, (4)Tohoku University, (5) National University of Singapore, (6) University of Copenhagen, (7) Takenote.ai, (8) University of Michigan, (9) Indian Institute of Technology Delhi, (10) Institute of Science Tokyo

   Submitted to ICLR 2026

   Abs PDF

   As Large Language Models (LLMs) gain stronger multilingual capabilities, their ability to handle culturally diverse entities becomes crucial. Prior work has shown that LLMs often favor Western-associated entities in Arabic, raising concerns about cultural fairness. Due to the lack of multilingual benchmarks, it remains unclear if such biases also manifest in different non-Western languages. In this paper, we introduce Camellia, a benchmark for measuring entity-centric cultural biases in nine Asian languages spanning six distinct Asian cultures. Camellia includes 19,530 entities manually annotated for association with the specific Asian or Western culture, as well as 2,173 naturally occurring masked contexts for entities derived from social media posts. Using Camellia, we evaluate cultural biases in four recent multilingual LLM families across various tasks such as cultural context adaptation, sentiment association, and entity extractive QA. Our analyses show a struggle by LLMs at cultural adaptation in all Asian languages, with performance differing across models developed in regions with varying access to culturally-relevant data. We further observe that different LLM families hold their distinct biases, differing in how they associate cultures with particular sentiments. Lastly, we find that LLMs struggle with context understanding in Asian languages, creating performance gaps between cultures in entity extraction.

2024

1. IGI

   Charting the Ethical Course: Navigating AI Advancements in Communication Education

   Huong Tra Le Nguyen*, Trung Tran*

   In Sanae Elmoudden, Jason S. Wrench (editors) The Role of Generative AI in the Communication Classroom

   Book
2. IGI

   Generative AI in terms of Its ethical problems for both teachers and learners: Striking a balance

   Huong Tra Le Nguyen*, Trung Tran*

   In Shalin Hai-Jew (editors) Generative AI in Teaching and Learning

   Book

2023

1. IGI

   Chatbots as motivational agents: Chatbots – The Value of a Digital Tool in Pedagogy

   Huong Tra Le Nguyen*, Trung Tran*

   In Mohammad Amin et al (editors) Trends, Applications, and Challenges of Chatbot Technology

   Book

2021

1. ICICT

   MAGNeto: An Efficient Deep Learning Method for the Extractive Tags Summarization Problem

   Hieu Trong Phung (1 and 2), Anh Tuan Vu (1), Tung Dinh Nguyen (1), Lam Thanh Do (1 and 2), Giang Nam Ngo (1), Trung Thanh Tran (1), Ngoc C. Lê (1 and 2) ((1) PIXTA Vietnam, Hanoi, Vietnam. (2) Hanoi University of Science and Technology, Ha Noi, Viet Nam.)

   The Seventh International Congress on Information and Communication Technology (ICICT), 2021

   Abs PDF

   In this work, we study a new image annotation task named Extractive Tags Summarization (ETS). The goal is to extract important tags from the context lying in an image and its corresponding tags. We adjust some state-of-the-art deep learning models to utilize both visual and textual information. Our proposed solution consists of different widely used blocks like convolutional and self-attention layers, together with a novel idea of combining auxiliary loss functions and the gating mechanism to glue and elevate these fundamental components and form a unified architecture. Besides, we introduce a loss function that aims to reduce the imbalance of the training data and a simple but effective data augmentation technique dedicated to alleviates the effect of outliers on the final results. Last but not least, we explore an unsupervised pre-training strategy to further boost the performance of the model by making use of the abundant amount of available unlabeled data. Our model shows the good results as 90% F1 score on the public NUS-WIDE benchmark, and 50% F1 score on a noisy largescale real-world private dataset. Source code for reproducing the experiments is publicly available at: https://github.com/pixta-dev/labteam.

2018

1. Symposium

   An enhanced X-Vector Model for Noise-Robust, Text-Independent Speaker Identification”. Panasonic Technology Symposium 2018

   Truong, T.A, Tran, T.T, Vu, D.L

   Abs PDF

   Our goal is todevelop an in-house, low-cost, scalable solution for closed-set text-independent speaker identification. Driven by Panasonic Corporation’s product development pathway, this research will be applicable a wide range of consumer and business-oriented products. Potential applications include recommendation systems for smart TVs, biometric authentication, and automatic transcription for conference meetings, among others. We propose three directions to enhance identification accuracy and noise-robustness: 1) retraining model on large open-source, noisy datasets; 2) omitting excessive data augmentation; 3) anew algorithm to detect non-voice samples. Our enhanced model achieves an increase in identification accuracy by 9% compared to existing state-of-the-art open-source solutions, and achieves non-voice detection accuracy of 97%. We are incorporating the speaker identification solution into a product prototype targeted at the Vietnamese B2B and B2C market.

2014

1. ComNavi-14

   Monitoring scintillations effects over Vietnam by means of a GNSS software receiver

   Workshop Communications and Navigations for the Development of Vietnam’s Marine Economy 2014

   Tran, T.T, Romero, R, Dovis, F

   Learning representations of data is an important problem in statistics and machine learning. While the origin of learning representations can be traced back to factor analysis and multidimensional scaling in statistics, it has become a central theme in deep learning with important applications in computer vision and computational neuroscience. In this article, we review recent advances in learning representations from a statistical perspective. In particular, we review the following two themes: (a) unsupervised learning of vector representations and (b) learning of both vector and matrix representations.

2013

1. IEEE

   Recent results in receiving and decoding signals from the Beidou system

   Duc Minh Truong*; Trung Thanh Tran*; Thuan Dinh Nguyen; Tung Hai Ta

   2013 International Conference on Localization and GNSS (ICL-GNSS)

   Abs PDF

   Since December 27, 2012, the Beidou Navigation Satellite System officially started to operate. This event is a great opportunity for researchers in South East of Asia to receive and analyze the Beidou signals. After the official statement, the researchers at NAVIS centre monitored the broadcasted signal by using NAVISOFT- our Software Radio Receiver. This paper shows the analysis on the navigation message that was broadcasted by the Beidou satellites on the B1I bandwidth. In general, we were able to observe a valid ephemeris data on visible satellites. The successful PVT computation by using combinations of GEO and MEO/IGSO in static condition through code-phase measurements is indicated in this paper.