Dr. Md Musfique Anwar

JU CIDL (Computational Intelligence, Data and Language) Research Lab is founded by Dr. Md Musfique Anwar, Professor, CSE dept, Jahangirnagar University. The lab focuses on various research domain such as Natual Langauge Processing, Artificical Intelligence, Online Social Network analysis, Machine Learning etc. Currently the lab members / researchers on this lab are actively contributing to the research world on various domain.

Research Project / Major Publication

BANGLASOCIALBENCH: A Benchmark for Evaluating Sociopragmatic and Cultural Alignment of LLMs in Bangladeshi Social Interaction

Large Language Models have demonstrated strong multilingual fluency, yet fluency alone does not guarantee socially appropriate language use. In high-context languages, communicative competence requires sensitivity to social hierarchy, relational roles, and interactional norms that are encoded directly in everyday language. Bangla exemplifies this challenge through its three-tiered pronominal system, kinship-based addressing, and culturally embedded social customs. We introduce BANGLASOCIALBENCH, the first benchmark designed to evaluate sociopragmatic competence in Bangla through context-dependent language use rather than factual recall. The benchmark spans three domains: Bangla Address Terms, Kinship Reasoning, and Social Customs, comprising 1,719 culturally grounded instances written and verified by native Bangla speakers. We evaluate twelve contemporary LLMs in a zero-shot setting and observe systematic patterns of cultural misalignment. Models frequently default to overly formal address forms, fail to recognize multiple socially acceptable address pronouns, and conflate kinship terminology across religious contexts. Our findings show that sociopragmatic failures are often structured and nonrandom; for example, inappropriate addressing choices concentrate heavily in downwardhierarchy (Elder→Younger) and informal contexts. This reveals persistent limitations in how current LLMs infer and apply culturally appropriate language use in realistic Bangladeshi social interactions.

Members: Tanvir Ahmed Sijan, S. M Golam Rifat (RUET), Pankaj Chowdhury Partha (BUET), Md. Tanjeed Islam

Publication: Acccepted as oral presentation (2nd highest average reviewer rating) at student research workshop at The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

AI-Enabled Deep Depression Detection and Evaluation Informed by DSM-5-TR

Depression, a prevalent and debilitating mental health disorder, affects millions of individuals worldwide, profoundly impacting their quality of life. Early identification and diagnosis of depression and its intensity are crucial for effective treatment and management. However, many people with depression do not seek professional help, especially in the early stages. In recent years, social media platforms like Twitter have gained popularity as spaces for sharing personal thoughts and emotions, including sensitive signals indicative of serious issues such as self-harm, suicidal thoughts, or illegal activities. These signals may help us to identify depression-related tweets and determine whether an individual is suffering from depression. This research focuses on utilizing artificial intelligence and deep learning (DL) models to categorize tweets related to depression and measure its intensity. The proposed approach combines emotional features, topical events, and behavioral-biometric signals to train the long short-term memory (LSTM)-based DL models. To create a comprehensive dataset, we collaborated with an expert psychologist who followed the clinical assessment procedure outlined in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5-TR) to label 95 322 tweets as “non-depressed” or “depressed,” further categorized into “mild,” “moderate,” and “severe” intensity levels. Through a series of experiments, our proposed method achieved superior performance compared to baseline models, yielding a mean squared error of 0.002336 and the highest R2 value of 0.61. These results highlight the accuracy and potential applications of our approach in automatic depression screening and monitoring on social media platforms.

Members: Tabia Tanzin Prama, Md. Saiful Islam (University of Newcastle, Australia)

Publication: IEEE Transactions on Computational Social Systems. 2024 (Journal Rank: Q1, SCI, IF: 5.0).

DOI: https://ieeexplore.ieee.org/document/10510456

BanglaNet: A Bangla Lip-reading System using Residual Network and Bi-directional Gated Recurrent Units

Visual speech recognition, often known as lip reading, seeks to identify the content of video utterances by analyzing lip movements. Recent years have seen several promising advancements, greatly aided by the recently released large-scale datasets for lipreading and the rapidly developing deep-learning algorithms. Most existing lip reading models are for the major languages. Unfortunately, no renowned audio-visual dataset is found for Bangla, the fifth-largest spoken language in the world. The most recent literature focuses on training lip reading models using a combination of frontend and backend networks. Most of these studies focus on the effectiveness of these networks. They try various combinations of these networks to improve lip reading accuracy. Also, most currently used approaches achieve outstanding performance by building a complicated neural network with several unique training strategies. But these techniques are always discussed in short. The proper application of these techniques could result in remarkable advancements without significantly altering the model. In our study, we developed the Bangla aUdio-visual Lipreading Indices (BULI) that consists of 500 most common Bangla words and phrases totalling ~25000 video samples from ~51 individuals. We conduct a thorough quantitative study and comparative analysis to demonstrate the effects of various options for lip reading. We consider the non-negligible impact of these techniques and the problematic state of developing an effective lip reading model for Bangla. We developed a network, “BanglaNet,” which consists of 3D CNN concatenated with ResNet18 as the front end and Bi-GRUs as the back end. On the BULI dataset, we compared our model to two cutting-edge baseline models. We achieve a top-5 accuracy of 92.34% for sentences and a top-5 accuracy of 72.1% for words by merely adding a few simple changes to the baseline pipeline.

Members: MD. Masum Bhuiyan

Award: 2nd Runner-up award in AI for Bangla 2.0 competition, 2023, Bangladesh.

Detecting sentiment dynamics and clusters of Twitter users for trending topics in COVID-19 pandemic

COVID-19 caused a significant public health crisis worldwide and triggered some other issues such as economic crisis, job cuts, mental anxiety, etc. This pandemic plies across the world and involves many people not only through the infection but also agitation, stress, fret, fear, repugnance, and poignancy. During this time, social media involvement and interaction increase dynamically and share one’s viewpoint and aspects under those mentioned health crises. From user-generated content on social media, we can analyze the public’s thoughts and sentiments on health status, concerns, panic, and awareness related to COVID-19, which can ultimately assist in developing health intervention strategies and design effective campaigns based on public perceptions. In this work, we scrutinize the users’ sentiment in different time intervals to assist in trending topics in Twitter on the COVID-19 tweets dataset. We also find out the sentimental clusters from the sentiment categories. With the help of comprehensive sentiment dynamics, we investigate different experimental results that exhibit different multifariousness in social media engagement and communication in the pandemic period.

Members: Md Shoaib Ahmed, Tanjim Taharat Aurpa

Publication: PLoS ONE, Volume: 16, Issue: 8, 2021. (Journal Rank: Q1, SCI, IF: 3.24).

DOI: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0253300

Attribute Driven Temporal Active Online Community Search

Almost all of the existing approaches to determining online local community are typically deliberated like-minded users who have similar topical interests. However, such methodologies overlook the prospective temporality of users' interests as well as users' degree of topical activeness. As a result, the consequential communities might have extremely lower active users. This research investigates how online social users' behaviors and topical activeness vary over time and how these parameters can be employed in order to improve the quality of the detected local community. For a given input query, consisting a query node (user) and a set of attributes, this research intends to find densely-connected community in which community members are temporally similar in terms of their activities related to the query attributes. To address the proposed problem, we develop a temporal activity biased weight model which gives higher weight to users' recent activities and develop an algorithm to search an effective community. The effectiveness of the proposed methodology is justified using four benchmark datasets and compared with four other baseline methods. Experimental results demonstrate that our proposed framework yields better outcomes than the baseline methods for all four benchmark datasets.

Members: Badhan Chandra Das

Publication: IEEE Access, Volume 9, pp. 93976 - 93989

DOI: https://ieeexplore.ieee.org/document/9467368

Page updated

Report abuse