Events

On Breaking Down Complex Questions

Speaker: Tomer Wolfson

Location: 60 Fifth Avenue, Room 650
Videoconference link: https://nyu.zoom.us/j/4202785259

Date: Wednesday, July 19, 2023

Abstract: (In this talk, Tomer will cover two recent projects that were part of his Ph.D. program and internship at Allen AI.) The first work presents a new method for answering multi-hop questions by meta-reasoning across multiple chains-of-thought. We prompt a large language model (LLM) to combine evidence from different sampled chains and produce the answer, along with an explanation. Our experiments show our solution improves the model's accuracy in 7 different multihop question-answering benchmarks that require skills such as commonsense, composition, comparison, and fact-verification.
  The second work, which is in progress, builds upon our previous research on question decomposition. We use question decompositions to build a new benchmark for answering questions that are more natural and complex than existing tasks. Our study shows how question decompositions can help answer questions that involve multiple answers from various sources such as tables, lists, and text. We will conclude by discussing the potential application of our task in evaluating the consistency of question-answering models and in generating answer provenance.

Bio: Tomer Wolfson is a fifth year Ph.D. student at the Blavatnik school of computer science at Tel Aviv University, and a research intern at the Allen Institute for AI. He is advised by Prof. Jonathan Berant and Prof. Daniel Deutch.
  His work is at the intersection of Natural Language Processing and Data Management and his main focus is developing algorithms that can comprehend complex questions across various domains and can offer justification and clarification for their outputs. In addition to this, he is passionate about building new NLP datasets that require more advanced reasoning skills and assessing the reasoning capabilities that existing models can or cannot easily learn. He is currently supported by the Israeli PBC Scholarship for Outstanding PhD students in Data Science.