Kapil Thadani

• Who? Principal research scientist at Yahoo Research NYC
  PhD in computer science from Columbia University
  Into artificial intelligence, natural language processing and machine learning
  
• More: Curriculum vitae 
  LinkedIn
  Google Scholar
  

Research interests

• Deep learning: large language models, vision-language models, knowledge distillation
• Representation learning: semantic, cross-lingual and multimodal representations, transfer learning
• Continual learning: online learning, active learning, deep reinforcement learning
• Structured prediction: information extraction, parsing, summarization, alignment and generation

Affiliations

• Natural language processing research for Yahoo News, Finance and Sports
• Columbia University NLP group, machine learning group and the center for computational learning systems
• The Association for Computational Linguistics

Teaching

• EECS 6984: Deep Learning for Computer Vision, Speech and Language
  Columbia University, Fall 2018
• EECS 6984: Deep Learning for Computer Vision, Speech and Language
  Columbia University, Spring 2017

Papers

• SCOT: Self-Supervised Contrastive Pretraining for Zero-Shot Compositional Retrieval 
  Bhavin Jawade, Joao Soares, Kapil Thadani, Deen Dayal Mohan, Erfan Eshratifar, Jack Culpepper, Paloma de Juan, Srirangaraj Setlur and Venu Govindaraju
  In proceedings of WACV 2025 in Tucson, AZ
• Scalable Detection of Salient Entities in News Articles 
  Eliyar Asgarieh, Kapil Thadani and Neil O'Hare
  arXiv preprint, 2024
• Multilingual Taxonomic Web Page Classification through Ensemble Knowledge Distillation 
  Eric Ye, Xiao Bai, Neil O'Hare, Eliyar Asgarieh, Kapil Thadani, Francisco Perez-Sorrosal and Sujyothi Adiga
  In IEEE Transactions on Knowledge and Data Engineering, 2024
• Salient Object-Aware Background Generation using Text-Guided Diffusion Models    arXiv    Supplement    Code 
  Erfan Eshratifar, Joao Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku and Paloma de Juan
  In proceedings of the Workshop on Generative Models for Computer Vision at CVPR 2024 in Seattle, WA
• Unifying Margin-Based Softmax Losses in Face Recognition 
  Yang Zhang, Simao Herdade, Kapil Thadani, Eric Dodds, Jack Culpepper and Yueh-Ning Ku
  In proceedings of WACV 2023 in Waikoloa, HI
• Multilingual Taxonomic Web Page Classification for Contextual Targeting at Yahoo 
  Eric Ye, Xiao Bai, Neil O'Hare, Eliyar Asgarieh, Kapil Thadani, Francisco Perez-Sorrosal and Sujyothi Adiga
  In proceedings of KDD 2022 in Washington, D.C.
• Powering COVID-19 Community Q&A with Curated Side Information 
  Manisha Verma, Kapil Thadani and Shaunak Mishra
  In proceedings of the Workshop on Knowledge Injection in Neural Networks at CIKM 2021 in the Gold Coast, Australia
• Effective Few-Shot Classification with Transfer Learning 
  Aakriti Gupta, Kapil Thadani and Neil O'Hare
  In proceedings of COLING 2020 in Barcelona, Spain
• Learning to Create Better Ads: Generation and Ranking Approaches to Ad Creative Refinement    arXiv 
  Shaunak Mishra, Manisha Verma, Yichao Zhou, Kapil Thadani and Wei Wang
  In proceedings of CIKM 2020 in Galway, Ireland
• Phans, Stans and Cishets: Self-Presentation Effects on Content Propagation in Tumblr  
  Michael Miller Yoder, Qinlan Shen, Yansen Wang, Alex Coda, Yunseok Jang, Yale Song, Kapil Thadani and Carolyn Rosé
  In proceedings of WebSci 2020 in Southampton, UK
• Unsupervised Neologism Normalization using Embedding Space Mapping 
  Nasser Zalmout, Aasish Pappu and Kapil Thadani
  In proceedings of the Workshop on Noisy User-generated Text at EMNLP 2019 in Hong Kong, China
• Lightweight Multilingual Extraction and Linking    Code    Dataset 
  Aasish Pappu, Roi Blanco, Yashar Mehdad, Amanda Stent and Kapil Thadani
  In proceedings of WSDM 2017 in Cambridge, UK
• The Role of Discourse Units in Near-Extractive Summarization    Code    Dataset 
  Junyi Jessy Li, Kapil Thadani and Amanda Stent
  In proceedings of SIGDIAL 2016 in Los Angeles, California (Nominated for best paper)
• Extractive Summarization under Strict Length Constraints 
  Yashar Mehdad, Kapil Thadani, Dragomir Radev, Amanda Stent, Youssef Billawala and Karolina Buchner
  In proceedings of LREC 2016 in Portorož, Slovenia
• Predicting the Impact of Scientific Concepts using Full-Text Features 
  Kathleen McKeown, Hal Daumé III, Snigdha Chaturvedi, John Paparrizos, Kapil Thadani, Pablo Barrio, Or Biran, Suvarna Bothe, Michael Collins, Kenneth Fleischmann, Luis Gravano, Rahul Jha, Ben King, Kevin McInerney, Taesun Moon, Diarmuid O'Seaghdha, Dragomir Radev, Clay Templeton and Simone Teufel
  In the Journal of the American Society for Information Science and Technology, 2016
• Approximation Strategies for Multi-Structure Sentence Compression 
  Kapil Thadani
  In proceedings of ACL 2014 in Baltimore, Maryland
• Supervised Sentence Fusion with Single-Stage Inference    Slides 
  Kapil Thadani and Kathleen McKeown
  In proceedings of IJCNLP 2013 in Nagoya, Japan
• Cluster-Based Web Summarization 
  Yves Petinot, Kathleen McKeown and Kapil Thadani
  In proceedings of IJCNLP 2013 in Nagoya, Japan
• Sentence Compression with Joint Structural Inference    Slides    Code 
  Kapil Thadani and Kathleen McKeown
  In proceedings of CoNLL 2013 in Sofia, Bulgaria
• A Joint Phrasal and Dependency Model for Paraphrase Alignment    Dataset 
  Kapil Thadani, Scott Martin and Michael White
  In proceedings of COLING 2012 in Mumbai, India
• On-the-fly Topic Adaptation for YouTube Video Transcription 
  Kapil Thadani, Fadi Biadsy and Daniel M. Bikel
  In proceedings of Interspeech 2012 in Portland, Oregon
• Identifying Event Descriptions using Co-training with Online News Summaries 
  William Yang Wang, Kapil Thadani and Kathleen McKeown
  In proceedings of IJCNLP 2011 in Chiang-Mai, Thailand
• Towards Strict Sentence Intersection: Decoding and Evaluation Strategies    Slides 
  Kapil Thadani and Kathleen McKeown
  In proceedings of the Workshop on Monolingual Text-to-Text Generation at ACL-HLT 2011 in Portland, Oregon
• Optimal and Syntactically Informed Decoding for Monolingual Phrase-Based Alignment    Slides 
  Kapil Thadani and Kathleen McKeown
  In proceedings of ACL-HLT 2011 in Portland, Oregon
• A Hierarchical Model of Web Summaries 
  Yves Petinot, Kathleen McKeown and Kapil Thadani
  In proceedings of ACL-HLT 2011 in Portland, Oregon
• Time-Efficient Creation of an Accurate Sentence Fusion Corpus    Dataset 
  Kathleen McKeown, Sara Rosenthal, Kapil Thadani and Coleman Moore
  In proceedings of NAACL-HLT 2010 in Los Angeles, California
• Corpus Creation for New Genres: A Crowdsourced Approach to PP Attachment    Slides    Dataset 
  Mukund Jha, Jacob Andreas, Kapil Thadani, Sara Rosenthal and Kathleen McKeown
  In proceedings of the Workshop on Creating Speech and Text Language Data with Amazon's Mechanical Turk at NAACL-HLT 2010 in Los Angeles, California
• Towards Semi-Automated Annotation for Prepositional Phrase Attachment 
  Sara Rosenthal, William J. Lipovsky, Kathleen McKeown, Kapil Thadani and Jacob Andreas
  In proceedings of LREC 2010 in Valletta, Malta
• A Framework for Identifying Textual Redundancy    Slides 
  Kapil Thadani and Kathleen McKeown
  In proceedings of COLING 2008 in Manchester, UK
• Density Estimation under Independent Similarly Distributed Sampling Assumptions    Addendum 
  Tony Jebara, Yingbo Song and Kapil Thadani
  In proceedings of NIPS 2007 in Vancouver, Canada
• Spectral Clustering and Embedding with Hidden Markov Models 
  Tony Jebara, Yingbo Song and Kapil Thadani
  In proceedings of ECML 2007 in Warsaw, Poland

Patents

• Systems and Methods for Image Compositing via Machine Learning
  Bhavin Jawade, Amir Erfan Eshratifar, Kapil Thadani, Paloma de Juan, Joao V. B. Soares and Jack Culpepper
  US Patent filed Mar 2024
• Systems and Methods for Using AI to Facilitate Image Editing
  Kapil Thadani, Akshay Bahadur, Dipen Rughwani, Paloma de Juan and Joao V. B. Soares
  US Patent filed Mar 2024
• Method and System for Webpage Classification and Content Delivery 
  Eric Ye, Xiao Bai, Neil O'Hare, Eliyar Asgarieh, Kapil Thadani, Francisco Perez-Sorrosal and Sujyothi Adiga
  US Patent filed Aug 2023
• System and Method for Generating Video in Target Language 
  Paloma de Juan, Alex J. Shaw, Eric M. Dodds, Benjamin J. Culpepper, Kapil Thadani, Lakshmi V. Kesiraju, Praveen Mareedu, Sanika Shirwadkar, Xingyue Zhou and Yueh-Ning Ku
  US Patent filed Jul 2022
• Generation and Presentation of Summary List based upon Article 
  Kapil Thadani, Philip Anthony Hairr, Lippe Oosterhof and Xi Gao
  US Patent filed Jun 2021
• Ranking User Comments on Media Using Reinforcement Learning Optimizing for Session Dwell Time 
  Kapil Thadani, Akshay Soni, Parikshit Shah, Troy Chevalier, Sreekanth Ramakrishnan, Aaron Nagao and Zhi Qu
  US Patent filed Jun 2019, granted Apr 2023
• Generating Presentations based on Articles 
  Arunkumar Balasubramanian, Kapil Thadani and Andrew Crews
  US Patent filed Dec 2018, granted Jan 2022
• Systems and Methods for Unsupervised Neologism Normalization of Electronic Content Using Embedding Space Mapping 
  Aasish Pappu, Kapil Thadani and Nasser Zalmout
  US Patent filed Oct 2018, granted Oct 2020
• Entity Disambiguation 
  Aasish Pappu, Roi Blanco, Yashar Mehdad, Amanda Stent and Kapil Thadani
  US Patent filed Feb 2017, granted Feb 2024
• Scalable and Effective Document Summarization Framework 
  Youssef Billawala, Yashar Mehdad, Dragomir Radev, Amanda Stent and Kapil Thadani
  US Patent filed Feb 2016, granted Oct 2020
• Speech Recognition with Topic-Specific Language Models 
  Daniel M. Bikel, Kapil Thadani, Fernando Pereira, Maria Shugrina and Fadi Biadsy
  US Patent filed Dec 2012, granted Apr 2016

Dissertations and other publications

• Multi-Structured Models for Transforming and Aligning Text    Code 
  Kapil Thadani
  PhD Dissertation, Columbia University, 2015
  Committee:  Kathy McKeown, Owen Rambow, Julia Hirschberg, Michael Collins, Hal Daumé III
• Decreasing Textual Redundancy
  Kapil Thadani
  Master's Thesis, Columbia University, 2007
  Committee:  Kathy McKeown, Owen Rambow, Julia Hirschberg
• Independent Similarly Distributed Sampling Assumptions for Semiparametric Density Estimation 
  Tony Jebara, Yingbo Song and Kapil Thadani
  In proceedings of the 2007 New York Academy of Sciences Symposium on Machine Learning in New York City

Datasets

• A collection of document IDs in the New York Times Annotated Corpus for which the corresponding summaries on the nytimes.com homepage are genuinely extractive or near-extractive. Code to extract these documents from the corpus is available here. 
  Download  (239 KB)    README       BibTeX 
• A corpus of 1020 phrase-based alignments derived from the Edinburgh paraphrase corpus including tokenization fixes, dependency graphs, named entity annotations and baseline alignments generated by METEOR. See Scott Martin's description for more details.
  Download  (1.6 MB)    README       BibTeX 
• A small corpus featuring 297 pairs of related newswire sentences, each with 10 fusions of varying correctness (5 intersections and 5 unions) generated by Mechanical Turk users.
  Download  (91 KB)      README       BibTeX 
• A collection of 941 prepositional phrase attachment cases over unstructured blog text. Candidates were chosen automatically and final judgments were made by humans responding to multiple-choice questions on Mechanical Turk.
  Download  (130 KB)    README       BibTeX 

Miscellany

• Candidacy exam on text-to-text generation
• Erdős number:  4   Me → { Tony Jebara → Tommi Jaakkola;  Kathy McKeown → Zvi Galil } → Noga Alon → Paul Erdős
  Bacon number:  ∞