INSCIT: Information-Seeking Conversations with Mixed-Initiative Interactions

TL;DR

This paper introduces INSCIT, a new dataset for studying mixed-initiative information-seeking conversations, highlighting the challenges in current models and providing a benchmark for future improvements.

Contribution

The creation of INSCIT, a comprehensive dataset with human-human interactions involving mixed-initiative responses over Wikipedia, supporting new research tasks and evaluation protocols.

Findings

Current models underperform compared to humans.

The dataset enables evaluation of evidence identification and response generation.

Results indicate significant room for improvement in conversational knowledge systems.

Abstract

In an information-seeking conversation, a user may ask questions that are under-specified or unanswerable. An ideal agent would interact by initiating different response types according to the available knowledge sources. However, most current studies either fail to or artificially incorporate such agent-side initiative. This work presents InSCIt, a dataset for Information-Seeking Conversations with mixed-initiative Interactions. It contains 4.7K user-agent turns from 805 human-human conversations where the agent searches over Wikipedia and either directly answers, asks for clarification, or provides relevant information to address user queries. The data supports two subtasks, evidence passage identification and response generation, as well as a human evaluation protocol to assess model performance. We report results of two systems based on state-of-the-art models of conversational knowledge identification and open-domain question answering. Both systems significantly underperform humans, suggesting ample room for improvement in future studies.