Sumanta Kashyapi

Clustering Optimization as Blackbox (COB)

This project presents a new technique to directly optimize for RAND index, a discrete clustering metric. To propagate meaningful gradients from discrete loss function based on RAND index, we use Blackbox backpropagation.

Read the paper: Learn The Big Picture: Representation Learning for Clustering

Context-Aware Triamese Similarity (CATS)

CATS similarity metric is designed to calculate pairwise similarity score between embeddings of text passages based on the current query-context.

Passage Similarity metric using Transformer Embeddings (PaSTE)

An ensemble method combining traditional text similarity metrics and state-of-the-art Transformer-based language understanding models, suitable for subtopic clustering of text passages.

TREC 2020 Podcast challenge

A deep neural model combining abstractive and extractive summarization, suitable for generating short summaries from audio transcripts for podcast users.