INSCIT: Information-Seeking Conversations with Mixed-Initiative Interactions
This work addresses the challenge of mixed-initiative interactions in conversational AI for information-seeking tasks, but it is incremental as it primarily provides a new dataset and benchmarks.
The authors tackled the problem of information-seeking conversations where user queries are often under-specified or unanswerable, by introducing InSCIt, a dataset of 4.7K turns from 805 human-human conversations, and found that state-of-the-art models significantly underperform humans, with systems based on conversational knowledge identification and open-domain QA showing ample room for improvement.
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.