Kartikeya Badola

Software Engineer at Google Research India. B.Tech from Indian Institute of Technology, Delhi, advised by Dr. Mausam and Dr. Parag Singla. Previously, an intern at MILA

Email  /  Google Research Webpage  /  Google Scholar  /  Twitter

Research

A study on continual learning from a multilingual perspective. Proposes finetuning pipelines using parameter-efficient finetuning methods to minimize language-specific catastrophic forgetting while encouraging positive cross-lingual transfer in this setup.

A simple BERT based model for multi-instance multi-label setting of distantly supervsied relation extraction (DS-RE). Outperforms exsisting state-of-the-art models on monolingual and multilingual DS-RE datasets.

A high quality dataset for multilingual distantly supervised relation extraction (DS-RE) along with the first baseline results for the task.

Use of a global channel with an inherent bottleneck to model communication between specialist modules in a modular neural network (as opposed to all-pair attention update). Achieves better generalization across tasks from various domains.

Service