Sivaji Bandyopadhyay

Debjit Dhar, Soham Lahiri, Tapabrata Mondal, Sivaji Bandyopadhyay. Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025). 2025.

Mathematical formulas are commonly used to demonstrate theories and basic fundamentals in the Science, Technology, Engineering, and Mathematics (STEM) domain. The burgeoning research in the STEM domain results in the mass production of scientific documents that contain both textual and mathematical terms. In scientific information, the definition of mathematical formulas is expressed through context and symbolic structure that adheres to strong domain-specific notions. Whereas the retrieval of textual information is well-researched, and numerous text-based search engines are present. However, textual information retrieval systems are inadequate for searching scientific information containing mathematical formulas, including simple symbols to complicated mathematical structures. The retrieval of mathematical information is in its infancy, and it requires the inclusion of new technologies and tools to promote the retrieval of scientific information and the management of digital libraries. This article provides a comprehensive study of mathematical information retrieval and highlights their challenges and future opportunities.

Multilingualism in India is widespread due to its long history of foreign acquaintances. This leads to the presence of an audience familiar with conversing using more than one language. Additionally, due to the social media boom, the usage of multiple languages to communicate has become extensive. Hence, the need for a translation system that can serve the novice and monolingual user is the need of the hour. Such translation systems can be developed by methods such as statistical machine translation and neural machine translation, where each approach has its advantages as well as disadvantages. In addition, the parallel corpus needed to build a translation system, with code-mixed data, is not readily available. In the present work, we present two translation frameworks that can leverage the individual advantages of these pre-existing approaches by building an ensemble model that takes a consensus of the final outputs of the preceding approaches and generates the target output. The developed models were used for translating English-Bengali code-mixed data (written in Roman script) into their equivalent monolingual Bengali instances. A code-mixed to monolingual parallel corpus was also developed to train the preceding systems. Empirical results show improved BLEU and TER scores of 17.23 and 53.18 and 19.12 and 51.29, respectively, for the developed frameworks.

A COVID-19 news covers subtopics like infections, deaths, the economy, jobs, and more. The proposed method generates a news summary based on the subtopics of a reader’s interest. It extracts a centroid having the lexical pattern of the sentences on those subtopics by the frequently used words in them. The centroid is then used as a query in the vector space model (VSM) for sentence classification and extraction, producing a query focused summarization (QFS) of the documents. Three approaches, TF-IDF, word vector averaging, and auto-encoder are experimented to generate sentence embedding that are used in VSM. These embeddings are ranked depending on their similarities with the query embedding. A Novel approach has been introduced to find the value for the similarity parameter using a supervised technique to classify the sentences. Finally, the performance of the method has been assessed in two different ways. All the sentences of the dataset are considered together in the first assessment and in the second, each document wise group of sentences is considered separately using fivefold cross-validation. The proposed method has achieved a minimum of 0.60 to a maximum of 0.63 mean F1 scores with the three sentence encoding approaches on the test dataset.

Knowledge representation is of immense importance in the field of artificial intelligence and natural language processing. The representation of knowledge goes hand in hand with automated reasoning as one of the key goals of representing knowledge effectively is being able to reason about it. Researchers of knowledge representation and reasoning have built techniques and methods that are the main source of development in computer science and have made tremendous progress in a wide variety of real-life applications, ranging from natural language processing to robotics and software engineering. Further research is required in order to allow a more active role in guiding the reasoning process through the knowledge representation framework. This article has discussed knowledge representation and reasoning and analyzed the major challenges and new opportunities where novel knowledge representation and reasoning research have had a major impact.

Machine translation deals with automatic translation from one natural language to another. Neural machine translation is a widely accepted technique of the corpus-based machine translation approach. However, an adequate amount of training data is required, and there is a need for the domain-wise parallel corpus to improve translational performance that shows translational coverages in various domains. In this work, a domain-specific parallel corpus is prepared that includes different domain coverages, namely, Agriculture, Government Office, Judiciary, Social Media, Tourism, COVID-19, Sports, and Literature domains for low-resource English-Assamese pair translation. Moreover, we have tackled data scarcity and word-order divergence problems via data augmentation and prior alignment concept. Also, we have contributed Assamese pre-trained LM, Assamese word-embeddings by utilizing Assamese monolingual data, and a bilingual dictionary-based post-processing step to enhance transformer-based neural machine translation. We have achieved state-of-the-art results for both forward (English-to-Assamese) and backward (Assamese-to-English) directions of translation.

Diabetic Retinopathy (DR) is caused as a result of Diabetes Mellitus which causes development of various retinal abrasions in the human retina. These lesions cause hindrance in vision and in severe cases, DR can lead to blindness. DR is observed amongst 80% of patients who have been diagnosed from prolonged diabetes for a period of 10–15 years. The manual process of periodic DR diagnosis and detection for necessary treatment, is time consuming and unreliable due to unavailability of resources and expert opinion. Therefore, computerized diagnostic systems which use Deep Learning (DL) Convolutional Neural Network (CNN) architectures, are proposed to learn DR patterns from fundus images and identify the severity of the disease. This paper proposes a comprehensive model using 26 state-of-the-art DL networks to assess and evaluate their performance, and which contribute for deep feature extraction and image classification of DR fundus images. In the proposed model, ResNet50 has shown highest overfitting in comparison to Inception V3, which has shown lowest overfitting when trained using the Kaggle’s EyePACS fundus image dataset. EfficientNetB4 is the most optimal, efficient and reliable DL algorithm in detection of DR, followed by InceptionResNetV2, NasNetLarge and DenseNet169. EfficientNetB4 has achieved a training accuracy of 99.37% and the highest validation accuracy of 79.11%. DenseNet201 has achieved the highest training accuracy of 99.58% and a validation accuracy of 76.80% which is less than the top-4 best performing models.