Misbah Al Mamun

Publications

Papers

Two peer-reviewed IEEE conference papers, four manuscripts under review, one project ongoing.

Peer-reviewed

2

Beyond Sentiment: The Influence of Game Pricing on Player Feedback in Steam ReviewsPublished

Misbah Al Mamun, Akash Datta Roni
2025 International Conference on Computer and Information Technology (ICCIT), 2025

Hierarchical Context-Aware Network (HC-Net): A Multi-Task Learning Approach for Robust MRI Disease DiagnosisPublished

Akash Datta Roni, Misbah Al Mamun, Kazi Mohammad Shahed, Abul Kasim Mohammad Abu Musa
2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence and Networking (QPAIN), CUET, 2026

Under review

4

Swara boundary detection in Indian classical fluteUnder review

A systematic comparison of five acoustic feature representations — MFCC, Log-Mel, pitch, RMS and CQT — for automatic swara boundary segmentation with an attention-guided CRNN, trained on synthetic melodies and evaluated on real recordings. Attention-pattern analysis explains why the features differ.

SylEmotions: emotion detection in SylhetiUnder review

A dual-script Sylheti emotion corpus in both Syloti Nagri and Bangla script, covering seven classes, benchmarked across classical, character-level and pretrained-transformer models to contrast native-script against transliterated performance.

Equity-aware parliamentary speech attributionUnder review

Comparative fine-tuning of DistilBERT under five class-imbalance-mitigation strategies for party attribution on the UK Hansard corpus, 1979–2018.

Bangla misogyny detectionUnder review

Full fine-tuning benchmarked against LoRA-based parameter-efficient fine-tuning on BanglaBERT, XLM-RoBERTa and Qwen2.5-1.5B for fine-grained misogyny categorisation on the OMANISHA benchmark, with SHAP-based interpretability and error analysis.

In progress

1

Bangla–Sylheti Nagri machine translationOngoing

A three-tier hybrid pipeline — exact match, then RAG-grounded generation, then flagging — with ByT5-small fine-tuning on a self-built parallel corpus. The work argues for byte-level models on language pairs whose script the model has never seen.