Journal Articles
Permanent URI for this collectionhttps://idr.nitk.ac.in/handle/123456789/19884
Browse
8 results
Search Results
Item Long short-term memory network for learning sentences similarity using deep contextual embeddings(Springer Science and Business Media B.V., 2021) Meshram, S.; Anand Kumar, M.Semantic text similarity (STS) is a challenging issue for natural language processing due to linguistic expression variability and ambiguities. The degree of the likelihood between the two sentences is calculated by sentence similarity. It plays a prominent role in many applications like information retrieval (IR), plagiarism detection (PD), question answering platform and text paraphrasing, etc. Now, deep contextualised word representations became a better way for feature extraction in sentences. It has shown exciting experimental results from recent studies. In this paper, we propose a deep contextual long semantic textual similarity network. Deep contextual mechanisms for collecting high-level semantic knowledge is used in the LSTM network. Through implementing architecture in multiple datasets, we have demonstrated our model’s effectiveness. By applying architecture to various semantic similarity datasets, we showed the usefulness of our model’s on regression and classification dataset. Detailed experimentation and results show that the proposed deep contextual model performs better than the human annotation. © 2021, Bharati Vidyapeeth's Institute of Computer Applications and Management.Item Video summarization and captioning using dynamic mode decomposition for surveillance(Springer Science and Business Media B.V., 2021) Radarapu, R.; Gopal, A.S.S.; Nh, M.; Anand Kumar, M.Video surveillance has become a major tool in security maintenance. But analyzing in a playback version to detect any motion or any sort of movements might be tedious work because only for a short length of the video there would be any motion. There would be a lot of time wasted in analyzing the video and also it is impossible to always find the accurate frame where the transition has occurred. So there is a need in obtaining a summary video that captures any changes/motion. With the advancements in image processing using OpenCV and deep learning, video summarization is no longer an impossible work. Captions are generated for the summarized videos using an encoder–decoder captioning model. With the help of large, well-labeled video data sets like common objects in context, Microsoft video description, video captioning is a feasible task. Encoder–decoder models are used extensively to extract text from visual features with the arrival of long short term memory (LSTM). Attention mechanism has been widely used on decoder for the work of video captioning. Keyframes are obtained from very long videos using methods like dynamic mode decomposition, an algorithm in fluid dynamics, OpenCV’s absdiff(). We propose these tools for motion detection and video/image captioning for very long videos which are common in video surveillance. © 2021, Bharati Vidyapeeth's Institute of Computer Applications and Management.Item A transformer-based architecture for fake news classification(Springer, 2021) Mehta, D.; Dwivedi, A.; Patra, A.; Anand Kumar, M.In today’s post-truth world, the proliferation of propaganda and falsified news poses a deadly risk of misinforming the public on a variety of issues, either through traditional media or on social media. Information people acquire through these articles and posts tends to shape their world view and provides reasoning for choices they take in their day to day lives. Thus, fake news can definitely be a malicious force, having massive real-world consequences. In this paper, we focus on classifying fake news using models based on a natural language processing framework, Bidirectional Encoder Representations from Transformers, also known as BERT. We fine-tune BERT for specific domain datasets and also make use of human justification and metadata for added performance in our models. We determine that the deep-contextualizing nature of BERT is effective for this task and obtain significant improvement over binary classification, and minimal yet important improvement in six-label classification in comparison with previously explored models. © 2021, The Author(s), under exclusive licence to Springer-Verlag GmbH Austria, part of Springer Nature.Item Downscaled XCO2 Estimation Using Data Fusion and AI-Based Spatio-Temporal Models(Institute of Electrical and Electronics Engineers Inc., 2024) Pais, S.M.; Bhattacharjee, S.; Anand Kumar, M.; Chen, J.One of the well-known greenhouse gases (GHGs) produced by anthropogenic human activity is carbon dioxide (CO2). Understanding the carbon cycle and how negatively it affects the ecosystem requires analysis of the rise in CO2 concentration. This work aims to map CO2 concentration for the entire surface, making it useful for regional carbon cycle analysis. Here, column-averaged CO2 dry mole fraction, called XCO2, measured by the orbiting carbon observatory-2 (OCO-2) satellite, is used. Because of spectral interference by the clouds and aerosols, there are many missing footprints in the Level-2 swath of OCO-2, making it disruptive to understand any assessment related to the carbon cycle. The objective of this work is to predict 1 km2 XCO2 using data resampling and machine learning models. This work achieves a minimum mean absolute error (MAE) and root mean square error (RMSE) of 0.3990 and 0.8090 ppm, using the monthly models. © 2004-2012 IEEE.Item Forecasting of Fine-Grained SIF of OCO-2 Using Multi-source Data and AI-Based Techniques(Springer, 2025) Pais, S.M.; Bhattacharjee, S.; Anand Kumar, M.The emission of unused energy absorbed from the sunlight by plants and other photosynthetic organisms is known as solar-induced fluorescence (SIF). SIF is a direct proxy for the photosynthetic activity of the plants used to monitor drought, crop yield estimation, ecological processes, and carbon cycles. Comprehending the SIF dynamics beforehand helps gain an understanding of vegetation dynamics, carbon cycle, and crop phenology. This study explores the potential of using Orbiting Carbon Observatory-2 (OCO-2) SIF data for forecasting SIF at regional scales. The research utilizes machine learning models and data fusion to forecast the SIF data, by establishing relationships between observed SIF from past timestamps and the Enhanced Vegetation Index. The lasso regression achieves minimal error of RMSE 0.0355 Wm-2nm-1sr-1 and MAPE of 16.9093% for forecasting monthly SIF data. In contrast, the light gradient boosting machine model (LG) performs well for a larger non-linear dataset, i.e., seasonal models achieving a RMSE of 0.0389 Wm-2nm-1sr-1 and MAPE of 17.4895%, respectively. Karnataka and Maharastra, the two Indian states, are considered as the study areas for this work for a temporal window of 2017–2019. Fine-grained, uniformly distributed SIF forecasting provides valuable insights for understanding vegetation responses to environmental changes, optimizing agricultural practices, and developing climate change mitigation strategies. © Indian Society of Remote Sensing 2025.Item LeDoFAN: enhancing lengthy document fake news identification leveraging large language models and explainable window-based transformers with n-gram expulsion(Springer Science and Business Media Deutschland GmbH, 2025) LekshmiAmmal, H.R.; Anand Kumar, M.Nowadays, people use social media to gather everything around them and consider it their primary source of information. Moreover, people rely more on information disseminated through social media and news channels. The alarming concern is that as the amount of information increases, the amount of fake news or misinformation spread also increases through social media. Generally, fake news has few lines of data; when it comes to a document or an article, the amount of information or the size of the documents is high, and it needs to be appropriately trained to build a model. In this work, we have developed a model that identifies and classifies fake news, consisting of articles collected from social media websites and news pages trained using transformer-based architecture. We have introduced a novel window method for handling lengthy documents and an N-gram expulsion method for managing similar words for classifying the article as fake or real news. We achieved the state-of-the-art F1-score of 0.3492 on test data for the window-based N-gram expulsion method and got an F1-score improvement of 2.1% for long documents alone with this method. We also explored the large language models (LLMs), specifically TinyLlama, which could only achieve an F1-score of 0.2098, and with LLama for summarization of the document that achieved an F1-score of 0.3402 with N-gram expulsion. We have further explored the results using Explainable Artificial Intelligence (XAI) to know the reason behind the proposed model’s intuition. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2025.Item High-Spatial-Resolution Estimation of XCO2 Using a Stacked Ensemble Model(Multidisciplinary Digital Publishing Institute (MDPI), 2025) Pais, S.M.; Bhattacharjee, S.; Anand Kumar, M.; Balamurugan, V.; Chen, J.Highlights: What are the main findings? The study develops a customized stacked ensemble model that generalizes (Formula presented.) predictions across multiple country, such as Germany, France, and Japan. It produces gap-filled high-resolution monthly, seasonal, and yearly maps, highlighting vegetation dynamics and seasonal cycles. What is the implication of the main finding? The customized stacked ensemble model provides reliable cross-country (Formula presented.) predictions at 1 (Formula presented.) resolution, validated against TCCON and CAMS, supporting large-scale environmental monitoring. Seasonal and yearly analyses show vegetation dynamics and photosynthetic activity significantly influence (Formula presented.), enhancing the model’s adaptability for agriculture, different climate assessments, and future global mapping. One of the leading causes of climate change and global warming is the rise in carbon dioxide ((Formula presented.)) levels. For a precise assessment of (Formula presented.) ’s impact on the climate and the creation of successful mitigation methods, it is essential to comprehend its distribution by analyzing (Formula presented.) sources and sinks, which is a challenging task using sparsely available ground monitoring stations and airborne platforms. Therefore, the data retrieved by the Orbiting Carbon Observatory-2 (OCO-2) satellite can be useful due to its extensive spatial and temporal coverage. Sparse and missed retrievals in the satellite make it challenging to perform a thorough analysis. This work trains machine learning models using the Orbiting Carbon Observatory-2 (OCO-2) (Formula presented.) retrievals and auxiliary features to obtain a monthly, high-spatial-resolution, gap-filled (Formula presented.) concentration distribution. It uses a multi-source aggregated (MSD) dataset and the generalized stacked ensemble model to predict country-level high-resolution (1 (Formula presented.)) (Formula presented.). When evaluated with TCCON, this country-level model can achieve an RMSE of 1.42 ppm, a MAE of 0.84 ppm, and (Formula presented.) of 0.90. © 2025 by the authors.Item A reasoning based explainable multimodal fake news detection for low resource language using large language models and transformers(Springer Nature, 2025) LekshmiAmmal, H.R.; Anand Kumar, M.Nowadays, individuals rely predominantly on online social media platforms, news feeds, websites, and news aggregator applications to acquire recent news stories. This trend has resulted in an increase in the number of available social media platforms, online news feeds, and news aggregator applications. These news platforms have been accused of spreading fake news to gain more attention and recognition. Earlier, this misinformation or fake news used to be propagated only in the text form. However, with the advent of technology, now it is spread in multimodal forms, such as images with text, videos, and audio with textual content. Currently, the automatic fake news detection models are focused on high resource languages and superficial output. Social media users need clarity and reasoning when it comes to identifying fake news, rather than just a superficial classification of news as fake. Providing context, reasoning, and explanations can help users understand why certain news is misleading or false. Hence, a multimodal system has to be developed to identify and justify fake news. In this proposed work, we have developed a multimodal fake news system for the Low Resource Language Tamil with reasoning-based explainability. The dataset for this proposed work is retrieved from fact-check websites and official news websites. We have experimented with different combinations of models for visual and text modalities. Further, we integrated LLM-based image descriptions into our model with the text and visual features, resulting in an F1 score of 0.8736. We used the Siamese model to determine the similarity of the news and its image descriptions. Additionally, we conducted error analysis and used explainable artificial intelligence to explore the reasoning behind our model’s predictions. We also present the textual reasoning for the model’s predictions and match them with images. © The Author(s) 2025.