I am currently working as an Applied Scientist at Amazon India. Prior to that, I was a Postdoctoral Researcher at Google Research India. I completed my PhD from the Department of Computer Science and Automation, Indian Institute of Science Bangalore under the guidance of Dr. Partha Pratim Talukdar. My thesis topic was on Knowledge Graph Embedding methods. Prior to that, I have been a master's student in the same department and worked under the guidance of Dr. Shivani Agarwal in the area of Machine Learning at the Machine Learning and Learning Theory Lab.
I am broadly interested in methods for Knowledge Graph (KG) creation and expansion and application of such background KGs for end tasks such as Question Answering and Document Classification. My current research focuses on improving KG completion methods and their application on densifying Open KGs extracted from a text corpus.
Working on Product Information Extraction.
Working on Language Models.
Worked on Search Query Recommendation.
Worked on Task Specific Knowledge Graph Construction.
Worked on conversation based searches on entertainment domain. It requires solving multiple sub-problems like named-entity recognition, user-intent detection etc. My work is focused on finding user intents and context management during conversation. I am also working on a template-based method which can be used for named-entity recognition and user-intent detection. It can also be used for generating suggestions as user types the query.
Coordinate Descent Algorithms form a class of simple optimization algorithms which has received attention of many researchers in last decade. There has been significant advancements in adapting these algorithms in parallel (multi-core) settings. In this project, we focused on studying parallel versions of Coordinate Descent Algorithms. We also implemented and conducted experiments with some of these algorithms.
Entity linking(EL) is a process of mapping textual mentions of named-entities in text to an entity in some knowledge base. EL is used in numerous areas of natural language processing to automate structured information retrieval from raw corpus. In this project, we focused on D2W (Disambiguation to Wikipedia) task, where we map textual mentions to corresponding Wikipedia pages. Specifically, we studied the effects of co-reference resolution (using Stanford CoreNLP) on the performance of Wikifier system for D2W task.
This project aims to develop a technique for estimating the probability of patients' mortality in the Indian intensive care units. We apply different machine learning techniques (specifically, linear and non-linear logistic regression) to this problem. We also propose a boosting-style approach for predicting patient mortality rates, which automatically builds a score-based system for Indian patient data.
Null derefence is a common bug in programs. This project applies the abstract interpretation framework for the analysis of null dereferences in Java programs using Soot framework.
This project automates the question paper generation process for examinations. It covers the process of creation of questions database, selection of questions for exams meeting certain criteria and generation of encrypted paper and its decryption.
The aim of the project was to implement a basic viewer which can render tetrahedral meshes read from a file. It also supports rendering of individual meshes and group of meshes at different scaling levels.