Multiple Question Generation from Presentation Transcripts (MQG)

Introduction

Preparing for an oral presentation is a common task in various domains, particularly in professional settings. For instance, researchers who have had their papers accepted at conferences need to deliver either an oral or poster presentation to share their findings with fellow researchers. Politicians must prepare for debates during election periods, while company managers are required to deliver speeches to update investors on company operations. When crafting their presentation drafts, a fundamental concern arises: what kinds of questions might the audience ask? We plan to explore the ability of LLMs from this aspect. We prepared a dataset called MQG [1], which contains the questions that professional analysts asked after listening to managers' presentations in the earnings conference call. 

Both automatic evaluation and human evaluation will be included in the final assessment. 

[1] Juan, Yining, et al. "Generating Multiple Questions from Presentation Transcripts: A Pilot Study on Earnings Conference Calls." Proceedings of the 16th International Natural Language Generation Conference. 2023.

Dataset

Each instance contains "presentation" and "questions". Our goal is to generate "questions" based on the given "presentation" We will use ROUGE-L for auto-evaluation, and the participants will evaluate other teams' system outputs manually. All evaluation records will be shared for future research. The guidelines for manual evaluation will be shared later. 

Important Dates: Time zone: Anywhere On Earth (AOE)

Policies

Shared Task Organizers