publications

2023

1. 

   Classifying Bangla Health Misinformation from Social Media Using Machine Learning

   Parinda Rahman, Emon Sarker, Mahima Ahsan, and 2 more authors

   In 2023 International Conference on Machine Learning and Cybernetics (ICMLC), 2023

   Abs Video

   The growing use of social media platforms offers a means to spread health-related information at an increased rate. However, disseminating health information on social media may be hazardous as the information often has no regulation. Therefore, it is crucial to find ways to classify health misinformation. Using machine learning models, this project utilizes mined Bangla text data and its English Translations to classify health misinformation from mined social media data. Multiple models were used such as Random Forest, XGBoost, SVC, etc. The highest reported precision in Bangla text and the English translation is 77% and 79% respectively. XGB classifier had the highest accuracy of 75% for English translation and the Extra Trees Classifier had the highest accuracy of 72% for Bangla text.

2. 

   BORO-LILABOTI - Building of Robust Orders for Leveraging Isolated Letter Accumulations By Ordering Teacher Insights for Bangla Handwriting Recognition

   T. Shahjahan, M. Ahsan, M. Islam, and 3 more authors

   2023

   Working Paper

   Abs

   One of the key aspects of artificial intelligence is that it enables human interaction through vision and language. As a result, we have many useful applications, such as Optical Character Recognition (OCR). Many industries have successfully fabricated OCR for printed texts with high accuracy. Most of the time, language is resourceful and does not cause morphological difficulties. OCR performs appallingly in recognizing morphologically rich and resource-limited texts, but it shows improvement in the context of handwritten text recognition. Any morphologically rich language, such as Bangla, Hindi, Arabic, Hebrew, etc., encounters class imbalance issues because of the language composition. Also, low-resource languages have less data available for training, which leads to a biased, non-convergent model. To address this issue, particularly in the Bangla language, we introduce an enhanced knowledge distillation-based approach: Unlocking Bias Mitigation in Bangla Handwritten Word Recognition Through Inter-Linguistic Character-Level Teacher Model Insights (BORO LILA-BOTI). This approach mitigates the bias toward major classes significantly. Our approach comprises a Convolutional Recurrent Neural Network (CRNN) student model that gains inter-linguistic insights from languages with similar character shape patterns. To evaluate the effectiveness of our model on unseen data, we perform inter-dataset training and testing on Bangla Writing and BN-HTRD. Our evaluations achieved a 3.96% and 8.35% increase in the F1-micro score for minor classes in Bangla Writing and BN-HTRD, respectively, compared to the base (No KD) model. It also surpasses the state-of-the-art f1 micro-score for the minor classes, which is Super teacher LILA-BOTI.

3. 

   FPS does increase performance: Video Action Recognition with Structured State Spaces

   M. Ahsan, E. Sarker, P. Rahman, and 3 more authors

   2023

   Working Paper

   Abs