How CoChat is Using the Human Factor to Improve Customer-Service AI

This article is part of the Academic Alibaba series and is taken from the paper entitled "CoChat: Enabling Bot and Human Collaboration for Task Completion" by Xufang Luo, Zijia Lin, Yunhong Wang, and Zaiqing Nie, first published in 2018 by the Association for the Advancement of Artificial Intelligence. The full paper can be read here.

Use of chat bots as online service tools is growing, but human agents continue to prove indispensable to customer satisfaction in complex cases. Faced with this reality, most approaches have tended to separate human and AI labor, introducing human agents to solve problems only when bots fail to deliver. By and large missing has been research on real-time human-AI collaboration models, with humans and bots working in a complementary manner rather than in isolation.

Now, a team of researchers from Alibaba AI Labs, Microsoft Research, and Beihang University have developed CoChat, an original dialogue management framework that enables chat bots to take new prompts from their human colleagues in real time while processing customer requests. As well as streamlining human labor, this gradually reduces the need for human intervention over time in a growing array of scenarios.

CoChat's developers call its ability to learn from human colleagues in real time "online learning". Previous systems have used a two-part learning model consisting of "supervised learning," where the chat bot learns from historic chat logs between the human worker and the user, and "reinforcement learning," where the bot learns from interactions with users. The addition of "online learning" to this existing combination is what makes CoChat unique.

 The CoChat Framework

CoChat's online learning model is made possible through a new type of hierarchical recurrent neural network (HRNN). The developers added an external memory to CoChat's HRNN, calling the new model a memory-enhanced HRNN, or "MemHRNN". When human workers input responses to newly encountered situations that the bots struggle to deal with, the bot's dialogue manager stores the responses in the external memory. It then attempts to apply these new suggestions to future scenarios, at which point human workers again interact with the bots as needed.

CoChat’s MemHRNN model applied to a restaurant booking task

CoChat's developers tested the bot on two real-life tasks: making a restaurant reservation and booking movie tickets. The tests concluded that CoChat was able to reduce human workloads by 91.35% in restaurant reservation and 89.32% in movie ticket booking. Furthermore, the learning-enhanced dialogue manager achieved impressive user satisfaction rates of 97.04% and 92.62% respectively for the two tasks.

These results show that CoChat's MemHRNN framework can effectively handle one-shot learning challenges when new actions are needed to address new user requests. As a practical dialogue management tool, CoChat breaks new ground by leveraging external memory and human-AI hybrid solutions to better satisfy user needs.

The full paper can be read here.

.     .     .