Publication and Under Review

Natural Language Processing

1. Md. Shahidul Salim, Sk Imran Hossain, An Applied Statistics dataset for human vs AI-generated answer classification, Data in Brief, 2024, 110240, ISSN 2352-3409,https://doi.org/10.1016/j.dib.2024.110240. 🌐paper 🔗Github

   🎯 We have developed a dataset of 4231 question-and-answer pairs for Applied Statistics, which includes 116 questions selected by domain experts and answers provided by 100 students and 50 AI-generated responses using ChatGPT. Our dataset can be used to develop AI detectors for the Applied Statistics domain and to evaluate existing ones. Additionally, our data collection framework can be applied to other domains.



2. Bose, D., & Salim, M. S. (2024). Suggesting Bengali words using masked language model. In 3rd international conference on computing advancements (ICCA).

   🎯 This study explores Bangla word suggestion in texts using a fine-tuned Bangla-BERT model with Masked Language Modeling (MLM). While transformer-based models like BERT have revolutionized natural language processing (NLP), the availability of task-specific datasets, especially for word-suggesting in Bangla, is limited. This research addresses this gap by employing MLM to predict Bangla words. Sub-word tokenization enhances the model's ability to handle unknown words. Unlike existing pretrained BanglaBERT model that uses word-piece masking, we experiment with dynamic whole-word masking algorithms. Our research aims to analyze the performance of the fine-tuned BanglaBERT model under different masking strategies and compare results to the traditional BanglaBERT approach.



3. Asif Mohammed Saad, Umme Niraj Mahi, Md. Shahidul Salim, Sk Imran Hossain, Bangla news article dataset, Data in Brief, 2024, 110874, ISSN 2352-3409, https://doi.org/10.1016/j.dib.2024.110874.

   🎯 In this research, we present an updated standard Bangla dataset based on gathered Bangla news articles. In total, more than 1.9 million articles from nine Bangla news websites were gathered; the selection process was led by a number of categories, including sports, economy, politics, local news, tech, tourism, entertainment, education, health, the arts, and many more. The dataset per newspaper contains varying attributes, such as title, content, time, tags, meta, category, etc. This dataset will enable data scientists to investigate and assess theories related to Bangla natural language processing. Furthermore, there is a greater chance that the dataset will be utilized for domain-specific large language models in the context of Bangladesh, and it may be used to develop deep learning and machine learning models that categorize articles according to subjects.



4. BCoQA: Benchmark and Resources for Bangla Context-based Conversational Question Answering(Submitted to EMNLP 2024(Revision completed)) 🌐OpenReview

   🎯 Developing a Bangla Context-based Conversational Question Answering (CCQA) system faces challenges like limited domain-specific data, poor translation methods, and a lack of pretrained language models. To tackle these issues, this work constructs a robust Bangla CCQA dataset using quality-controlled machine translation and large language model-based augmentation of existing English CCQA datasets. The dataset is then divided into training, validation, and test splits. Various sequence-to-sequence models are fine-tuned and evaluated using the train and test splits, with conversation history included in the input prompts to maintain context.



5. Nabil, A., Das, d., Salim, M. S., Arifeen, S., & Fattah, H. M. A. (2023). Bangla emergency post classification on social media using transformer based bert models. In 6th international conference on electrical information and communication technology (EICT 2023). (Accepted) 🌐paper

   🎯 Text classification is one of the most important tasks in Natural Language Processing. As text data is growing rapidly, it needs more computational power to classify the text in a big dataset. The task is difficult for characteristic-rich languages like Bangla. Having good-quality text data significantly affects the outcome of the model that has been used to classify them. Nowadays, social media can be an important source of information. But there is a huge number of data which are of no use. As the use of social media is increasing day by day, people are posting about events around their surroundings. So, it can be an important propaganda of the media. In this study, various text classification methods were used to classify the texts in Bangla from social media, which can be categorized as emergencies that may need immediate actions from the government, local authority, or law enforcement or even may need international attention. Therefore, 5839 social media posts were collected from Facebook and Twitter, which were written in Bangla along with some mixed English words. Then, after preprocessing, various Machine Learning models, Deep Neural Network models, and Transformer based models were applied to classify them. Among these models, transformer-based XLM-RoBERTa outperformed all the other models with an F1-score of 95.25.



6. Salim, M. S., Murad, H., Das, D., & Ahmed, F. (2023). "BanglaGPT: A Generative Pretrained Transformer-Based Model for Bangla Language," 2023 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD), Dhaka, Bangladesh, 2023, pp. 56-59, doi: 10.1109/ICICT4SD59951.2023.10303383.🌐paper

   🎯 Natural Language Processing (NLP) has entered a new era with the advent of pre-trained language models, paving the way for constructing robust language models. Pretrained transformer-based models such as GPT-2 have become prevalent due to their cutting-edge efficiency. However, these approaches rely heavily on resource-intensive languages, forcing other languages to adopt multilingual frameworks (mGPT). The mGPT model could perform better for low-resource languages such as Bangla because the model has been trained on a diverse dataset spanning multiple languages. Recent studies show that the language-specific GPT model outperforms the multilingual mGPT model. In this research, we have proposed a pretrained monolingual GPT model called BanglaGPT using the objective of causal language modeling (CLM). Due to the lack of available large datasets for NLP tasks in Bangla, we have created a Bangla language model dataset called BanglaCLM using a 26.24 GB Bangla corpus scraped from several public websites. We have used a subword-based tokenization algorithm named Byte-Pair Encoding (BPE) for Bangla and finally trained the Bangla-GPT2 model from scratch using the BanglaCLM dataset. Our pretrained BanglaGPT provides state-of-the-art performance for Bangla text generation with a perplexity score of 2.86 and a loss score of 0.45 on the test set.



7. S. Salim, T. Islam, R. Zannat, N. Mia, M. Fuad and H. Murad, "Towards Developing a Transformer-Based Bangla Typing Error Correction Model: A Deep Learning-Based Approach," 2023 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD), Dhaka, Bangladesh, 2023, pp. 75-78, doi: 10.1109/ICICT4SD59951.2023.10303361.🌐paper

   🎯 Despite being the sixth most widely spoken language globally, the Bangla language faces a lack of effective and efficient automated methods to correct keyboard errors in written text. Such errors can result in communication misunderstandings and misinterpretations, particularly in the digital domain, where written text plays a crucial role in communication and is heavily relied upon. However, no keyboard error model for the Bangla language has been found due to the lack of available datasets. In this research work, we have proposed a probabilistic approach to generate a synthetic Bangla keyboard typing error correction (BKTEC) dataset. By utilizing the novel BKTEC dataset, we have fine-tuned a transformer-based model for Keyboard error correction. After training it to convergence, the proposed model was tested on a test dataset, and we discovered a word level accuracy of 88.56%.



8. T. Ahmed, S. Hossain, M. S. Salim, A. Anjum and K. M. Azharul Hasan, "Gold Dataset for the Evaluation of Bangla Stemmer," 2021 5th International Conference on Electrical Information and Communication Technology (EICT), Khulna, Bangladesh, 2021, pp. 1-6, doi:10.1109/EICT54103.2021.9733662.🌐paper

   🎯 Stemming is a preprocessing task for Natural Language Processing(NLP) that involves normalizing inflected words representing the same concept of the original word. Steaming is a process of text normalization that transforms a slated word into its root form. It has a great impact in different applications in NLP and Natural Language Understanding. One of the biggest challenges in Bangla stemming is to collect the rules that are associated with Bangla word stemming and standard dataset for testing the accuracy. This paper presents a summery of rules associated with Bangla stemming and a gold standard dataset corpus which will helps to testify the stemming algorithm. We verify the rules and using our rules and corpus we tested with existing Bangla stemming algorithms and found that the proposed corpus makes a difference with existing techniques.



9. M. S. Salim Shakib, T. Ahmed and K. M. Azharul Hasan, "Designing a Bangla Stemmer using rule based approach," 2019 International Conference on Bangla Speech and Language Processing (ICBSLP), Sylhet, Bangladesh, 2019, pp. 1-4, doi: 10.1109/ICBSLP47725.2019.201533.🌐paper 🔗Github

   🎯 Stemming is a preprocessing task for natural language processing that involves normalizing inflected words representing the same concept of the original word. Steaming is a process of text normalization that has many applications. There are many techniques for steaming of inflected words for different languages but very few works for Bangla word steaming. Therefore, stemming Bangla word is a unsolved problem. There are many different situations that can occur in Bangla language for word steaming. In this paper, we present a rule based algorithm to stem Bangla words. We developed the rules for infection detection for verb inflection, number inflection, and others. Using our rules, we developed a system to find the root word of Bangla words and found good performance. Sufficient examples are provided to explain the proposed system.

Machine Learning/Deep Learning

1. "Trisha, Shahid, Md. Shahidul Salim , Jeba and Mahbub, "Automated Classification of Gastrointestinal Polyps from Endoscopic Images Using a Deep Learning Approach" 2024 International Conference on Recent Progresses in Science, Engineering and Technology (Accepted)

   🎯 Research suggests that approximately 20-30% of adults aged 50 and older have colorectal polyps, a prevalent type of gastrointestinal (GI) polyps detected during colonoscopy exams. An increased risk of cancer development is associated with certain polyps. Timely identification is essential for implementing preventative measures and preventing the advancement of polyps to cancer. Considering the importance of early diagnosis, this study was conducted to classify GI polyps using a deep-learning approach. The Gastrovision dataset was used for this study. The ResNet50 model pre-trained on ImageNet was employed for transfer learning, achieving training, validation, and testing accuracies of 91.99%, 92.29%, and 89%, respectively. The specificity, precision, and F1-score obtained were 98.97%, 97.16%, and 81.65%, respectively. A user interface has been built to assist users in selecting and identifying images as part of a decision support system for categorizing GI polyps from endoscopic images.



2. R. T. H. Promi, R. A. Nazri, M. S. Salim and S. M. T. U. Raju, "A Deep Learning Approach for Non-Invasive Hypertension Classification from PPG Signal," 2023 International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM), Gazipur, Bangladesh, 2023, pp. 1-5, doi: 10.1109/NCIM59001.2023.10212940.🌐paper

   🎯 Hypertension or high blood pressure is the leading risk factor with the largest contribution to the burden of disease and mortality. Therefore, regular blood pressure (BP) checks are necessary to track hypertension and reduce risk. In light of this, we have proposed a noninvasive method for hypertension detection based on photoplethysmography (PPG) signals using deep learning techniques. A hybrid deep learning model consisting of a Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) is introduced. First, PPG signals were acquired from 219 subjects, which underwent pre-processing steps. Then, hypertension was detected with satisfactory results by using our proposed model. Here, CNN is used to extract features in order to work better in classification tasks. Our proposed model constructed with CNN and GRU achieved an Accuracy of 85.00 %, Recall of 84.44 %, Precision of 87.77% and F1-score of 86.08%. Our goal is to improve the medical industry by precisely detecting hypertension.



3. Hossain, L., Hossain, I., Salim, M. S., Raju, S. M. T. U., & Saha, J. (2023). A novel technique for classification of motor imagery EEG signal based on deep learning approaches. In Proceedings of the 2nd international conference on big data, IoT and machine learning (bim 2023). (Accepted)🌐paper

   🎯 Motor imagery classification is an important area of research that aims to provide insights into the neural processes underlying motor control. By analyzing the patterns of brain activity related to imagined movement, researchers can better understand how the brain originates and controls voluntary movements. This knowledge can have significant implications for the development of novel therapies for neurological diseases that impact motor performance, such as stroke and Parkinson's disease. In this study, the performance of three deep learning methods, including Artificial Neural Networks (ANNs), Deep Neural Networks (DNNs), and Bidirectional Long Short-Term Memory (Bi-LSTM), were evaluated for the classification of motor imagery EEG data. The results of the study showed that DNNs and Bi-LSTM performed better than the other algorithms with Bi-LSTM obtaining the highest average accuracy 96.84%, sensitivity 95%, specificity 98.72%, F1-score 96.80%. This finding highlights the potential of these methods for EEG-based BCI applications. Moreover, a detailed comparison of several deep learning methods for the classification of motor imagery was presented, providing useful insights for future research toward the development of efficient and accurate BCIs. In conclusion, this study demonstrated the effectiveness of DNNs and Bi-LSTM for the classification of motor imagery EEG data and their potential for use in the development of BCIs for neurological diseases. These findings can guide future research in the field and contribute to the development of more efficient and accurate BCIs.



4. Ashiqussalehin, M., Jahan, K., Rahaman, M., & Salim, M. (2022). Human Abnormal Behavior Detection Using Convolution Neural Network. Specialusis Ugdymas, 1(43), 4076–4083.🌐paper

   🎯 When a behavior is regarded as unusual in particular circumstances, it is termed abnormal. The term "abnormal behavior" has many definitions depending on the situation. People sprinting in a field, for example, is regarded natural, but it appears weird if it occurs in a mall. Similarly, loitering in alleyways, arguing, or pushing each other in public places are all regarded weird in certain situations. Deep learning has been widely used in the computer vision sector in recent years, with considerable success in human activity recognition. The proposed technique helps in the detection of abnormal behavior under a variety of contexts, including background changes, the number of subjects (one, two people, or a crowd), and a variety of unusual human behaviors.

Under Review

1. LLM based QA chatbot builder: A generative AI-based chatbot for question answering(Submitted to softwareX)
📝This software describes the development of a web application called the LLM QA Builder, designed to streamline the creation of LLM-based interactive chatbots for organizational information retrieval. The application integrates various development phases including data collection and preprocessing, LLM fine-tuning, testing, inference, and chat interface development. It supports fine-tuning multiple LLMs such as Zephyr, Mistral, Llama-3, Phi, Flan-T5, and user-provided models, with enhanced retrieval capabilities via retrieval-augmented generation (RAG). It also includes an automatic web crawling RAG data scraper and a human evaluation feature for model quality assessment. The system's capabilities are demonstrated through a university information chatbot, with comparative analysis of different LLMs using a benchmark crowd-sourced dataset.

2. Agricultural Recommendation System based on Multivariate Weather Forecasting Model(Submitted to Annals of Data Science) 🌐PRE-PRINT 🔗Github
📝This paper proposes a context-based crop recommendation system using a weather forecast model to improve farming practices in Bangladesh. The multivariate Stacked Bi-LSTM Network is used for accurate weather prediction, including rainfall, temperature, humidity, and sunshine. The system guides farmers in making informed decisions about planting, irrigation, harvesting, and more. It also alerts farmers about extreme weather conditions and provides knowledge-based crop recommendations for flood and drought-prone areas.

3. A Suffix Independent Algorithm for Stemming Bangla Words using Finite State Transducer
📝This study proposes and evaluates a suffix-independent stemming algorithm for Bangla language using a finite state transducer (FST)-based framework. The algorithm creates a dictionary of root words implemented as an FST, achieving high speed and no memory usage for vocabulary keeping. A novel stemmer dataset was developed to evaluate the algorithm's performance, resulting in 96.58% detection accuracy and 96.34% stemming accuracy. The proposed scheme outperforms existing methods and demonstrates effectiveness through experiments.

4. Detecting AI-Generated Assignments in Educational Evaluation: A Transformer-Based Approach
📝This research work presents a transformer-based model to detect whether an assignment is AI-generated or human-written. The model was trained on a dataset of 5410 assignments, with 2742 being AI-generated and 2668 being human-written. Among the explored transformer-based architectures, DistilBERT provided the highest accuracy of 92%.

Ongoing Research

My projects