DEEP-DIAL 2018

Report on AAAI 2018 Workshop on Reasoning and Learning for Human-Machine Dialogues (DEEP-DIAL18; Photos)

Natural conversation has been a key sub-area of AI for decades. Their most recent form, chatbots, which can engage people in natural conversation and are easy to build in software, have been in the news a lot lately. There is a mad rush by companies to release chatbots to show their AI capabilities and gain market valuation. However, beyond basic demonstration, there is little experience in how they can be designed and used for real-world applications that need decision making under constraints (e.g., sequential decision making). Furthermore, there is an upcoming interest and need for innovation in Human-Technology-Interaction, as addressed in the context of Companion Technology and Social Robots. Here, the aim is to implement technical systems that smartly adapt their functionality to their users’ individual needs and requirements and are even able to solve problems in close co-operation with human users. To this end, they need to enter into a dialog and convincingly explain their suggestions and decision-making behavior.

Although, statistical and machine learning methods are well entrenched for language understanding and entity detection, however, the wider problem of dialog management is unaddressed with mainstream tools supporting rudimentary rule-based processing.

There is an urgent need to highlight the crucial role of reasoning methods such as constraints satisfaction, planning and scheduling, and learning working together with them, that can play to build an end-to-end conversation system that evolves over time. From the practical side, conversation systems need to be designed for working with people in a manner that they can explain their reasoning, convince humans about making choices among alternatives, and stand up to ethical standards demanded in real life settings.

To discuss these, the workshop brought together over 100 AI researchers from around the world to discuss a bouquet of research topics around human-machine dialogs. The program included 4 invited talks, 7 reviewed full paper presentations and 4 lightening talks accompanied by posters, and a topical panel discussion.

The day started with an invited talk by Paul Crook of Facebook titled “Statistical Machine Learning for Dialog Management; its history and future promise”. Paul argued that since there is no single type of conversation, a single approach to handle all classes of dialogs does not look possible. He explained the four approaches for building dialog management systems: finite-state based, frame-based, inference based and response-generation based; and delved on their relative trade-offs. The second invited talk was by Erik Mueller of Capital One who talked about Eno, a chatbot deployed at a major bank in his talk titled, “Continuous Improvement of Intelligent Assistants Through Interaction with Live Customers”. In the third invited talk, titled “Combining Functional Programming, Probabilistic Reasoning, and Machine Learning in an Event- Driven AI Agent Framework”, Matt Davis of IBM Research talked about an agent framework called CHIA to develop dialog systems using building blocks of services, actions, skills and agents. The framework has been used to quickly develop chatbots for a slew of applications while being scalable and additive to new capabilities over time. The final invited talk was by Kristiina Jokinen of Helsinki University and NAIST on “Interactions with Social Robots - issues and challenges in combining knowledge and dialogue capabilities for digital companions”. Her work was in the context of social robots and contrasted with other invited talks where agents were virtual. Further, they had to take humans physical state into account while planning and interacting with them.

The program had authors of peer-reviewed papers discussing ideas ranging from using neural networks, knowledge graphs and discourse trees to generate conversation, to their applications in scheduling meetings, financial industry and learning when to speak. They generated a lot of questions and discussions. The day ended with an engaging panel moderated by Kartik Talamadupula of IBM Research and including all the invited speakers, on the topic – “Challenges in Wide-Scale Adoption of Chatbots to Help Those Needy (Children, Elderly or Physically Challenged)”. A few notable points that came out were that chatbots are equally relevant and important to society’s segments that are traditionally not targeted. Further, building a default behavioral assumption in a chatboat may create issues or missed opportunities for them. Example - preference for short conversation or using informal language. Panelists discussed on who are dis-advantaged and whether having a single agent for all v/s have a personalized agent for each person were desirable along with associated costs.

The event was an exciting day of technical exchange and attendees expressed views that

followup workshops will be worthwhile in order to build research momentum around a topic of significant application potential.

Concluding note:

Biplav Srivastava, Susanne Biundo, Ullas Nambiar and Imed Zitouni served as cochairs of the workshop. The papers of the workshop were published as AAAI Press Technical Report WS-18-14. More details are available at the workshop web site: http://www.zensar.com/deep-dial18