Selected work

Representative Papers

A Spatio-Temporal Deep Learning Approach For High-Resolution Gridded Monsoon Prediction

Jan 05, 2026arXiv.org

This study addresses the limitations of traditional monsoon forecasting, which typically provides coarse, regionally averaged seasonal point estimates that are insufficient for fine-scale management. For the first time, Indian summer monsoon prediction is formulated as a spatiotemporal video prediction task. A deep learning model based on convolutional neural networks is developed, using multivariate atmospheric and oceanic fields from January to May as multi-channel image sequences. Leveraging ERA5 reanalysis and India Meteorological Department (IMD) observational data, the model establishes a high-resolution gridded mapping from the precursor period to the monsoon season (June–September). This approach overcomes conventional constraints by enabling high-spatial-resolution forecasts of both monthly and total seasonal rainfall, thereby supporting refined climate outlooks at both intraseasonal and seasonal timescales.

0 citationsRead paper

LLMPR: A Novel LLM-Driven Transfer Learning based Petition Ranking Model

May 27, 2025

India’s judicial system suffers from severe case backlogs, and traditional manual prioritization of legal petitions is inefficient and highly subjective. This paper proposes the first automated petition prioritization framework integrating semantic representations—derived from legal-domain language models (e.g., LegalBERT)—with lightweight quantitative features (e.g., filing interval, character count). Crucially, we find that a simple model using only these numerical features achieves exceptional performance: R² = 0.988 and Spearman’s ρ = 0.998—matching or surpassing multimodal baselines (e.g., Random Forest and Decision Tree classifiers with >99% accuracy). This challenges the prevailing assumption that large language models are indispensable for such tasks. Our approach is rigorously validated on the ILDC dataset (7,593 manually annotated petitions), demonstrating substantial improvements in both timeliness and objectivity of judicial case triage.

0 citationsRead paper

NBF at SemEval-2025 Task 5: Light-Burst Attention Enhanced System for Multilingual Subject Recommendation

May 06, 2025

This work addresses cross-lingual topic classification for English–German academic literature. We propose a lightweight sentence embedding framework that employs token-level self-attention to compress internal dimensionality and integrates negative sampling with margin-based retrieval loss under a bilingual joint training paradigm, substantially reducing GPU memory and computational requirements. Compared to existing lightweight models, our approach achieves an average recall of 32.24% across all topics and qualitative evaluation scores of 43.16% and 31.53%, demonstrating significant performance gains. Our core contribution is the first coupling of token-granularity self-attention with margin loss for cross-lingual academic text embedding learning—achieving high retrieval effectiveness at extremely low computational cost. This establishes a scalable new paradigm for multilingual topic recommendation in resource-constrained settings.

0 citationsRead paper

Recent publications

Latest Papers

A Spatio-Temporal Deep Learning Approach For High-Resolution Gridded Monsoon Prediction

Jan 05, 2026arXiv.org

This study addresses the limitations of traditional monsoon forecasting, which typically provides coarse, regionally averaged seasonal point estimates that are insufficient for fine-scale management. For the first time, Indian summer monsoon prediction is formulated as a spatiotemporal video prediction task. A deep learning model based on convolutional neural networks is developed, using multivariate atmospheric and oceanic fields from January to May as multi-channel image sequences. Leveraging ERA5 reanalysis and India Meteorological Department (IMD) observational data, the model establishes a high-resolution gridded mapping from the precursor period to the monsoon season (June–September). This approach overcomes conventional constraints by enabling high-spatial-resolution forecasts of both monthly and total seasonal rainfall, thereby supporting refined climate outlooks at both intraseasonal and seasonal timescales.

0 citationsRead paper

LLMPR: A Novel LLM-Driven Transfer Learning based Petition Ranking Model

May 27, 2025

India’s judicial system suffers from severe case backlogs, and traditional manual prioritization of legal petitions is inefficient and highly subjective. This paper proposes the first automated petition prioritization framework integrating semantic representations—derived from legal-domain language models (e.g., LegalBERT)—with lightweight quantitative features (e.g., filing interval, character count). Crucially, we find that a simple model using only these numerical features achieves exceptional performance: R² = 0.988 and Spearman’s ρ = 0.998—matching or surpassing multimodal baselines (e.g., Random Forest and Decision Tree classifiers with >99% accuracy). This challenges the prevailing assumption that large language models are indispensable for such tasks. Our approach is rigorously validated on the ILDC dataset (7,593 manually annotated petitions), demonstrating substantial improvements in both timeliness and objectivity of judicial case triage.

0 citationsRead paper

NBF at SemEval-2025 Task 5: Light-Burst Attention Enhanced System for Multilingual Subject Recommendation

May 06, 2025

This work addresses cross-lingual topic classification for English–German academic literature. We propose a lightweight sentence embedding framework that employs token-level self-attention to compress internal dimensionality and integrates negative sampling with margin-based retrieval loss under a bilingual joint training paradigm, substantially reducing GPU memory and computational requirements. Compared to existing lightweight models, our approach achieves an average recall of 32.24% across all topics and qualitative evaluation scores of 43.16% and 31.53%, demonstrating significant performance gains. Our core contribution is the first coupling of token-granularity self-attention with margin loss for cross-lingual academic text embedding learning—achieving high retrieval effectiveness at extremely low computational cost. This establishes a scalable new paradigm for multilingual topic recommendation in resource-constrained settings.

0 citationsRead paper