Hi!
I'm a Research Engineer at Google DeepMind, where I work on improving code capabilities for Gemini and apply it to products.
I'm was previously a graduate student at the Language Technologies Department, Carnegie Mellon University, for the MIIS program. I was part of the MultiComp Lab, where I worked with Chaitanya Ahuja and Prof. Louis-Philippe Morency on multimodal gesture generation.
I interned at DeepMind for the summer (2022), working on code generation. I was previously a Research Engineer (on contract) at Google Research India, working with Dr. Partha Talukdar in the NLU Group. Here, I worked on entity-aware translation for educational content, and neural temporal parsing of queries.
Before this, I spent an amazing year at Microsoft Research India, and worked with Dr. Monojit Choudhury and Dr. Kalika Bali on various problems in low-resource language systems and probing NLI models.
I did my undergraduate thesis at Microsoft Research under Navin Goyal and Monojit Choudhury, on semantic parsing applied to the conversion of natural language to regular expressions and SQL. I graduated with a B.E. (Hons.) in Computer Science from BITS Pilani, Goa, India, in 2019.
Experience
-- Devised continual learning techniques for generative models. Focused on preventing catastrophic forgetting in multimodal gesture generation.
-- Work published at ICCV'23.
-- Trained and implemented large language models for code generation. (AlphaCode style models).
-- Explored multi-task training and reinforcement learning objectives for sampling and re-ranking.
-- Created a scalable entity-aware translation+transliteration pipeline to generate subtitles for English-medium college lecture videos (from NPTEL) for various Indian languages.
-- Developed neural temporal parsing models for multilingual queries using knowledge graph infused techniques.
-- Contributed to creating a challenge for Indian multilingual QA.
-- Probed large pretrained lanuage models for NLI reasoning, by designing a taxonomy of reasoning capabilities, annotating an existing NLI dataset based on the capabilities, then evaluating the model performance on this new dataset.
-- Worked on a range of multilingual problems, such as quantitative analysis of language diversity in ACL, efficacy of code-mixing chatbots, measuring quality of crowdsourced speech data, among others.
-- Works published at ACL, CoNLL, CSCW, LREC, EMNLP Workshop, and ICON (see publications below).
-- Worked on semantic parsing and its applications to natural language to code. Particularly focused on natural language to regular expressions (NL2Regex), and to SQL queries.
Implemented and experimented with different neural architectures for NL2Regex, and conducted an error analysis. Subsequently, I also analyzed quality of various semantic parsing datasets
using a range of metrics and indicators.
-- Research carried out also contributed to my undergraduate thesis.
-- Created voice interface application for truck drivers. Used pocketsphinx-android from CMUSphinx to power voice recognition.
Customized to recognize key commands in English, Hindi, and Indian-English.
-- Created capabilities like an in-built support system, location and route support, and verbal data entry.
-- Presented prototype to I-Loads administration (CEO,CFO,CTO) and tech team.
Publications
Gemini: A Family of Highly Capable Multimodal Models
Gemini Team Google:..., , ..., Jeffrey Dean, Oriol Vinyals
Arxiv 2023
website |
pdf |
abstract
Coverage/Media Mentions:
New York Times |
BBC |
Tech Crunch |
Wired Magazine
Continual Learning for Personalized Co-speech Gesture Generation
Chaitanya Ahuja, , Ryo Ishii, Louis-Philippe Morency
International Conference on Computer Vision (ICCV) 2023
website |
pdf |
code
TaxiNLI: Taking a Ride up the NLU Hill
, Somak Aditya*, Aalok Sathe*, Monojit Choudhury
The SIGNLL Conference on Computational Natural Language Learning 2020 (CoNLL) 2020
website |
pdf |
code
The State and Fate of Linguistic Diversity and Inclusion in the NLP World
, Sebastin Santy*, Amar Budhiraja*, Kalika Bali, Monojit Choudhury
Annual Conference of the Association for Computational Linguistics (ACL) 2020
website |
pdf |
abstract
Coverage/Media Mentions:
FTC Commissioner Speech |
NLP with Friends |
Quartz |
Ruder's Blog |
LacunaFund |
Underrated ML |
NLP Newsletter |
SIGTYP Newsletter
Do Multilingual Users Prefer Chat-bots that Code-mix? Let's Nudge and Find Out!
Anshul Bawa, Pranav Khadpe, , Kalika Bali and Monojit Choudhury
ACM Conference on Computer Supported Cooperative Work and Social Computing (CSCW) 2020
pdf
Crowdsourcing Speech Data for Low-Resource Languages from Low-Income Workers
Basil Abraham, Danish Goel, Divya Siddarth, Kalika Bali, Manu Chopra, Monojit Choudhury, , Preethi Jyoti, Sunayana Sitaram and Vivek Seshadri (Alphabetically Ordered)
International Conference on Language Resources and Evaluation (LREC) 2020
pdf
CoSSAT: Code-Switched Speech Annotation Tool
Sanket Shah, , Sebastin Santy, Sunayana Sitaram
AnnoNLP Workshop: Empirical Methods in Natural Language Processing (EMNLP) 2019
pdf |
abstract |
slides
Unsung Challenges of Building and Deploying Language Technologies for Low Resource Language Communities
, Christain Barnes, Sebastin Santy, Simran Khanuja, Sanket Shah, Anirudh Srinivasan, Satwik Bhattamishra, Sunayana Sitaram, Monojit Choudhury and Kalika Bali
International Conference on Natural Language Processing (ICON) 2019
pdf |
abstract |
slides
Coverage/Media Mentions:
Indian Express |
Microsoft Stories
Talks and Panels
The Nuts and Bolts (and Nuances) of Foundation Models
Centre for Democracy and Technology Panel, Remote
video |
description
NLP for the Long Tail
Google Research, Remote
The State and Fate of Linguistic Diversity in the NLP World
NLP with Friends, Remote
video |
description
Reviewing