Research Areas
Research Map
The diagram below shows how my research areas are connected, with Large Language Models (LLMs) as the central unifying theme.
Large Language Models (LLMs)
Large Language Models have fundamentally transformed how machines understand and generate natural language. My research investigates the capabilities and limitations of LLMs, with a particular focus on their performance in low-resource and multilingual settings. I explore how LLMs can be adapted, fine-tuned, and evaluated to handle linguistically diverse inputs — especially for Indic languages where training data is scarce. A key concern in my work is developing robust evaluation frameworks that go beyond surface-level metrics and align more closely with human judgement.
Natural Language Processing
My NLP research spans a broad range of tasks including sentiment analysis, emotion classification, question answering, and multilingual text understanding. A core theme across this work is low-resource NLP — developing methods that perform reliably even when labeled data is limited. I have worked extensively on code-mixed text (where speakers switch between languages within a sentence), as well as on building and benchmarking NLP systems for Indic languages such as Marathi and Hindi. This research aims to make NLP tools more equitable and accessible for speakers of underrepresented languages.
Knowledge Graphs
Knowledge Graphs (KGs) provide a structured representation of real-world entities and their relationships, forming the backbone of many modern AI systems. My research addresses the challenge of cross-lingual entity alignment — identifying equivalent entities across knowledge graphs built in different languages. I developed one of the first approaches for Hindi–English entity alignment using multilingual BERT and FastText embeddings, outperforming traditional graph-based methods. My ongoing work explores integrating KG reasoning with neural language models to enhance factual grounding and interpretability.
Explainable AI
As AI systems become more deeply embedded in high-stakes decision-making, the ability to understand and explain their outputs becomes critical. My research in Explainable AI (XAI) focuses on making NLP model decisions interpretable and transparent to both researchers and end users. I work on building datasets that include human-aligned explanations for model predictions, and on designing systems that produce rationales alongside their outputs. This work connects closely with my research on social bias detection, where explanations are essential for understanding the nature and source of harmful content.
Neuro-symbolic AI
Neuro-symbolic AI seeks to combine the pattern recognition power of neural networks with the structured reasoning capabilities of symbolic systems. My interest in this area stems from a desire to build NLP systems that are not only accurate but also logically consistent and interpretable. I explore how symbolic representations — such as meaning representations and knowledge graphs — can be integrated with neural architectures to improve reasoning over natural language, particularly for tasks that require multi-step inference or structured world knowledge.
Resource and Evaluation
Building high-quality resources and designing meaningful evaluation frameworks are foundational to progress in NLP. My work in this area includes creating annotated datasets for sentiment, emotion, social bias, and question answering — particularly for languages and domains that are underserved by existing benchmarks. I also investigate the limitations of standard automatic metrics and propose evaluation strategies that better reflect human perception and real-world utility. This is especially critical for low-resource Indic languages, where off-the-shelf metrics often fail to capture linguistic nuance.
AMR and UMR Parsing
Abstract Meaning Representation (AMR) and Uniform Meaning Representation (UMR) are graph-based formalisms that capture the deep semantic structure of natural language sentences. AMR abstracts away from surface form to represent the core meaning of a sentence as a rooted, directed graph, while UMR extends this to cross-lingual and cross-document settings. My research explores the development of parsers for these representations, with a focus on adapting them to low-resource languages. These parsers serve as a foundation for downstream tasks such as question answering, information extraction, and natural language reasoning.