Go to ACL ARR 2024 June homepageHandling Dialog Dependencies to Reformulate Requests in Human-Agent Interaction
Given the continual emergence of digital agents that employ tools and engines to satisfy multiple-nature user requests, there arises a critical need for efficiently orchestrating dialog in human-agent interactions. A fundamental function of this orchestration is to recognize user intent and send the appropriate request to the right engine/tool. However, given a dialog is conducted, information about the request might span through the whole conversation. In this work, we investigate the ability of large language models to recognize the user request in multi-turn human-agent interactions, considering dependencies in dialog and also reformulate it as a stand-alone sentence to be used for intent recognition and activation of tools, and engines without memory cells. To evaluate models as orchestrators, a demonstration dataset consisting of 42 dialogs, between an agent specialized in satellite data archives and a user, is developed and made publicly available. Thirteen models have been tested and five of them give outputs that comply with reference requests, with Gemini Pro 1.5 coming first.