Nikhil Mehta

Nikhil Mehta

6th Year PhD student at Purdue University

Publications

Interactively Learning Social Media Representations Improves News Source Factuality Detection

Published in Findings of IJCNLP-AACL 2023, 2023

We develop an interactive framework to detect the factuality of news sources. We take advantage of human insight, where humans interact on sub-graphs, providing us information about user similarity. This interaction is simple for humans and very quick. It also doesn’t require humans to be aware of the factuality of news content. We then combine this interaction with a graph deep learning framework, showing how these interactions lead to significant performance improvements. Our novel contribution is where humans should interact (for maximum impact), what they should interact on, and how we should incorporate the interactions.

Download here

An Interactive Framework for Profiling News Media Sources

Published in NAACL 2024, 2023

We develop a framework that combines the strengths of Large Language Models (LLMs), graphs, and humans to better profile news media sources (detect their factuality and political bias). Our framework performs better than using each approach independently: We outperform LLMs and graph models, and we need a lot fewer human interactions (less than 5) than having humans profile all news content.

Download here

Improving Grounded Language Understanding in a Collaborative Environment by Interacting with Agents Through Help Feedback

Published in Findings of EACL 2024, 2023

We develop an agent that can interact with humans to solve challenging grounded language learning tasks. Specifically, the agent can receive “help” from humans as feedback, and incorporate it to perform better in other scenarios. Further, the can detect when it is confused and ask a relevant clarification question, which a human (or the agent itself) can answer, allowing it to perform better.

Download here

Improving Natural Language Interaction with Robots Using Advice

Published in NAACL 2019, 2019

We focus on the blocks-world task and propose a framework in which advice, high-level observations about the task that can help constrain the agents prediction, can improve performance. The advice can be provided via natural language by humans, even after the system is trained.

Download here