One Agent To Rule Them All: Towards Multi-agent Conversational AI

Findings of the Association for Computational Linguistics: ACL 2022, pages 3258 - 3267 May 22-27, 2022 c⃝2022 Association for Computational Linguistics One Agent To Rule Them All: Towards Multi-agent Conversational AI Christopher Clarke1 Joseph J. Peper1 Karthik Krishnamurthy2 Walter Talamonti2 Kevin Leach3∗ Walter Lasecki∗ Yiping Kang1 Lingjia Tang1 Jason Mars1 1University of Michigan, Ann Arbor, MI 2Ford Motor Company, Dearborn, MI 3Vanderbilt University, Nashville, TN {csclarke, jpeper, ypkang, lingjia, profmars}@umich.edu {kkrish65, wtalamo1}@ford.com, kevin.leach@vanderbilt.edu, wslasecki@gmail.com Abstract The increasing volume of commercially avail- able conversational agents (CAs) on the mar- ket has resulted in users being burdened with learning and adopting multiple agents to ac- complish their tasks. Though prior work has explored supporting a multitude of domains within the design of a single agent, the in- teraction experience suffers due to the large action space of desired capabilities. To ad- dress these problems, we introduce a new task BBAI: Black-Box Agent Integration, focus- ing on combining the capabilities of multi- ple black-box CAs at scale. We explore two techniques: question agent pairing and ques- tion response pairing aimed at resolving this task. Leveraging these techniques, we design One For All (OFA), a scalable system that pro- vides a unified interface to interact with multi- ple CAs. Additionally, we introduce MARS: Multi Agent Response Selection, a new en- coder model for question response pairing that jointly encodes user question and agent re- sponse pairs. We demonstrate that OFA is able to automatically and accurately integrate an en- semble of commercially available CAs span- ning disparate domains. Specifically, using the MARS encoder we achieve the highest accu- racy on our BBAI task, outperforming strong baselines. 1 Introduction Influenced by the popularity of intelligent conver- sational agents (CAs), such as Apple Siri and Ama- zon Alexa, the conversational AI market is grow- ing at an increasingly rapid pace and is projected to reach a valuation of US $13.9 billion by 2025 (Mar- ket and Markets, 2020). These CAs have already begun to show great promise when deployed in domain-specific areas such as driver assistance (Lin et al., 2018), home automation (Luria et al., 2017), and food ordering (Frangoul, 2018) with platforms ∗Work was done while at University of Michigan Figure 1: An example interaction using One For All which integrates multiple production black-box agents into a unified experience. such as Pandora and Facebook today hosting more than 300,000 of these agents (Chaves and Gerosa, 2018; Nealon, 2018). Most CAs are designed to be specialized in a sin- gle or set of specific domains. As such, users are required to interact with multiple agents in order to complete their tasks and answer their queries as shown in figure 1. For example, a user may use an agent such as Amazon Alexa for online shopping but engage with Google Assistant for daily news updates. Additionally, a given agent may be more proficient at a specific domain over another i.e A finance CA is better suited to answer finance ques- tions. As a result, users are taxed …

Content truncated to 500 words