This work presents two new neural architectures—one based on bidirectional LSTMs and conditional random fields, and the other that constructs and labels segments using a transition-based approach inspired by shift-reduce parsers. The models learn character-level representations of words to alleivate the problem of hand-crafting features. The models obtain state-of-the-art performance in NER in four languages without resorting to any language-specific knowledge or resources such as gazetteers.

As part of my multimodal machine learning class, I created a dataset containing modality-specific (audio, visual and text) annotations of sentiment for 27 YouTube videos of human interviews. I also presented a baseline for multimodal sentiment analysis on this dataset using multimodal denoising autoencoders and multi-layer perceptrons.

This paper presents a framework that provides literature-based evidence for predicted protein-protein interactions (PPIs). It uses distribtuional semantics from verb attachment information to construct a feature representation for a protein. The representations of two proteins are compared to understand if they have the potential to interact. There paper can be found here.

I participated in the MediaEval 2013 Placing Task. The aim of this task was to determine the exact geographical coordinates (Latitude and Longitude) of an image obtained from Flickr. We design a textual metadata filtering approach to extract geographically significant Flickr tags to help in location estimation. The working notes paper presented at Barcelona, Spain can be found here.

I interned at an awesome startup in India called Invention Labs from December 2013 - August 2014. The startup develops AAC (Augmentative and Alternative) communication tools for kids. During this time I worked on a language engine called FreeSpeech that converts sentence semantics generated by interactively asking the user a bunch of questions into well formed language syntax. To understand the motivation behind such an application and to see it in action please visit the FreeSpeech website.

What started off as a class project with a professor from undergrad turned into a generic project to develop visualization libraries that can adapt to a variety of datasets. We developed python libraries for bargraphs, pie charts and star charts from scratch using the Pygame graphics library. Building the graphics from scratch gave us full control over the appeal and interactivity of the visualizations. We applied this library to the education domain, specifically university ranking data provided by U-Multirank. We presented the results at the European Conference on Education 013 at Brighton, UK. The proceedings can be found here.

My Bachelor's thesis under the guidance of Professor Satyajit Ghosh was an interdisciplinary project in the fields of Computer and Atmospheric science. We ran the Weather Research and Forecasting Model (WRF), an open source mesoscale climate model developed by eminent atmospheric and computer scientists. We used it to predict the possiblity of an imminent cyclone on the east cost of India and developed a robust and fully automated framework capable of alerting coastal dwellers in such an event. We published the details of this work in the Atmospheric Science Letters journal. The paper can be found here. We were on the Indian news channels for this piece of work Headlines Today news interview