Project MAIA: Multilingual AI Agent Assistant

Project MAIA: Multilingual AI Agent Assistant Jo˜ao Grac¸a1, Paulo Dimas1, Helena Moniz2, Andr´e F. T. Martins1,3, Graham Neubig4 1Unbabel / Lisbon, Portugal 2INESC-ID / Lisboa, Portugal 3Instituto de Telecomunicac¸˜oes / Portugal 4CMU / Pittsburgh, USA {joao,pdimas,andre.martins}@unbabel.com, helena.moniz@inesc-id.pt, gneubig@cs.cmu.edu Abstract This paper presents the Multilingual Artifi- cial Intelligence Agent Assistant (MAIA), a project led by Unbabel with the collab- oration of CMU, INESC-ID and IT Lis- bon. MAIA will employ cutting-edge ma- chine learning and natural language pro- cessing technologies to build multilingual AI agent assistants, eliminating language barriers. MAIA’s translation layer will em- power human agents to provide customer support in real-time, in any language, with human quality. 1 Introduction Online conversational support – chat – is the fastest growing customer service channel, being the pre- ferred way for millennials to obtain customer ser- vice. Today, supporting international customers in this channel is mostly done by using human agents that speak different languages – a scarce and costly resource. The tremendous progress of language technologies (machine translation and di- alogue systems) in the last years makes them an appealing tool for multilingual customer service. However, current systems are still too brittle and impractical: first, they require too much data and computing power, failing for domains or languages where labeled data is scarce; second, they do not capture contextual information (e.g. current MT systems work on a sentence-by-sentence basis, ig- noring the conversation context); third, fully au- tomatic systems lack human empathy and fail on unexpected scenarios, leading to low customer sat- c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. isfaction. In MAIA, we will develop a multilin- gual conversational platform where human agents are assisted by AI agents. This approach will over- come the above limitations by targeting the follow- ing scientific and technological goals: • New memory-efficient neural models for context-aware machine translation, suitable for online and real-time translation. These models will retain key aspects of a conversation (e.g., the gender of the customer), bringing them up whenever needed to translate a message. • New answer generation techniques where the human agent (e.g., a tourism officer) will receive suggestions that reduce effort and increase the customer’s (e.g. a tourist) satisfaction. • New techniques for conversational quality es- timation and sentiment analysis to assess how well the conversation is addressing the cus- tomer’s needs, while simultaneously increasing “human empathy”. • Integration of the scientific advances above into a full end-to-end product. To this end, two demonstrators will be built to cover concrete use cases in the Travel and Tourism industries. 2 Overview of MAIA Figure 1 displays a mock-up of the user interface to assist the human agent. Illustrated is the conversa- tion history (on the agent’s language), a list of an- swer suggestions, a message box supporting auto- completion where the agent can type the response, and an indicator of the sentiment of the customer throughout the conversation. The overarching goal of MAIA is to build context-aware, …

Content truncated to 500 words