Better embedding representation for passages, entities, types and relations
We are studying how to embed entities, types and relations to infer new edges in knowledge graphs. We are investigating how to represent passages to compare them.
Complex multi-modal question answering
With IBM Research, I am exploring how to translate complex queries involving knowledge base access, arithmetic and logical operations into structured programs with memory.
Code-switched text analysis
Indian languages borrow heavily from English, resulting in ``code switching'' languages like Hindlish, Benglish, etc., the lingua franca of social media. We are investigating how to improve standard NLP tasks by generating synthetic code-switched text, and designing multi-task low-supervision recurrent networks.
Searching the annotated Web with entities, types and relations
We built CSAW, a search system that integrates type and role annotations with keyword matches, thereby exploiting lexical ontologies and entity taggers.