Two of our research papers have been accepted at EMNLP 2025

We are thrilled to announce that two of our research papers — BanMime and DM-Codec — have been accepted at EMNLP 2025, one of the most prestigious conferences in NLP (Core Rank A*).

BanMime has been accepted into the EMNLP 2025 Main Conference

DM-Codec has been accepted into the EMNLP 2025 Findings

BanMime: A Novel Bangla Meme Dataset for Visual Metaphor Explanation

In this work, we introduce BanMime, the first-ever dataset for Bangla memes enriched with human-annotated explanations for metaphors.

We curated a dataset of 2,000 misogynistic memes in Bangla. Each meme was labeled, and based on those labels, explanations were annotated to uncover the underlying metaphor.

Our experiments reveal that existing open-source Vision-Language Models (VLMs) struggle to comprehend visual metaphors, and their performance is even worse when generating explanations.

This work is a joint collaboration between CUET, DIU, and CCDS.

Authors: Md Ayon Mia*, Akm Moshiur Rahman Mazumder*, Khadiza Sultana Sayma, Md Fahim, Md Tahmid Hasan Fuad, MUHAMMAD IBRAHIM KHAN, AKM MAHBUBUR RAHMAN

(*Equal Contribution)

DM-Codec: Distilling Multimodal Representations for Speech Tokenization

DM-Codec is a novel speech tokenizer that integrates multimodal (acoustic, semantic, and contextual) representations via a language model (LM) and self-supervised speech model (SM) through distillation.

DM-Codec outperforms state-of-the-art models, reducing Word Error Rate (WER) by up to 13.46% and enhancing speech intelligibility and quality on the LibriSpeech dataset.

This work is done in collaboration with AWS (GenAI), Qatar Computing Research Institute, and University of Virginia

Authors: Md Mubtasim Ahasan, Md Fahim, Tasnim Mohiuddin, A K M Mahbubur Rahman, Aman Chadha, Tariq Iqbal, M Ashraful Amin, Md Mofijul Islam, Amin Ahsan Ali

Congratulations to the all co-authors, collaborators and supervisors