Ariyan Hossain

Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing

Ariyan Hossain, K. K. Rabbi, Dr. Farig Yousuf Sadeque, Dr. S. M. Taiabul Haque

Under Review, ARR (targeting EMNLP 2026) (2026)

Conducted the first controlled study (N=153) examining how LLM writing suggestions propagate gender stereotypes during human-LLM co-authoring of career essays. Demonstrated bias transfer through a two-measure evaluation framework combining an agentic language gap (via a fine-tuned agency classifier) and a novel Stereotype Congruence Rate (SCR) metric.

Benchmarking Bengali Dialectal Bias: A Multi-Stage Framework Integrating RAG-Based Translation and Human-Augmented RLAIF

K. M. J. Sami, D. Sumit, Ariyan Hossain, Dr. Farig Yousuf Sadeque

Under Review, ARR (targeting EMNLP 2026) (2025)

Created a benchmark evaluating 19 LLMs across 4,000 gold-labeled question sets spanning nine Bengali dialects using RAG-based translation. Proposed a RLAIF bias evaluation framework with Chain-of-Thought rubrics and introduced the CBS (Cultural Bias Score) metric for safety-critical assessment.

A Comparative Analysis of RAG Techniques for Bengali Standard-to-Dialect Machine Translation Using LLMs

K. M. J. Sami, D. Sumit, Ariyan Hossain, Dr. Farig Yousuf Sadeque

BLP Workshop @ IJCNLP-AACL 2025 (Published) (2025)

Designed two RAG pipelines for standard-to-dialectal Bengali translation across six dialects using in-context learning. Reduced WER from 76% to 55% with sentence-pair RAG, enabling Llama-3.1-8B to outperform GPT-OSS-120B.

Identifying Framing Bias in Online Climate Change News Articles

Ariyan Hossain, Dr. Muhammad Nur Yanhaona

Built a frame-classified climate change news dataset via expert annotation and GPT-4 synthetic augmentation for multi-class framing classification. Fine-tuned Longformer, BigBird, BERT, and LLaMA-3.1 for long-document classification, applying QLoRA for memory-efficient parameter-efficient fine-tuning (PEFT).

Exploring and Mitigating Gender Bias in Encoder-Based Transformer Models

Ariyan Hossain, K. M. A. Hannan, R. Haque, N. T. Rafa, H. Musarrat, S. A. Dipu, Dr. Farig Yousuf Sadeque

Undergraduate Thesis (2025)

Proposed MALoR, a novel metric for quantifying gender bias via masked-token probabilities in encoder-based models. Created gender-balanced pretraining datasets using Counterfactual Data Augmentation (CDA) by swapping gendered pronouns and names, generating over 19,000 balanced sentence pairs. Achieved up to 95% reduction in gender bias (e.g., "he–she" bias in BERT-base decreased from 1.27 → 0.08 and "his–her" from 2.51 → 0.36 MALoR score).