1Building an effective dialog system

At Facebook AI Research (FAIR), understanding dialogue is an ambitious and long-term AI research goal.

A truly effective dialogue system will be an assistive technology that will likely include a chatbot-type of system being able to interact with people through natural language communication. This could help people better understand the world around them and communicate more effectively with others, effectively bridging communication gaps. Researching and developing these kinds of technologies will only become more important as the amount of digital content continues to grow.

The attempt to understand and interpret dialogue is not a new one. As far back as 20 years, there were several efforts to build a machine a person could talk to and teach how to have a conversation. These incorporated technology and engineering, but were single purposed with a very narrow focus, using pre-programmed scripted responses.

Thanks to progress in machine learning, particularly in the last few years, having AI agents being able to converse with people in natural language has become a more realistic endeavor that is garnering attention from both the research community and industry.

However, most of today’s dialogue systems continue to be scripted: their natural language understanding module may be based on machine learning, but what they execute or answer is in general dictated by if/then statements or rules engines. While they are improvement on what existed decades ago, it is in large part due to the large databases of content used to create and script their responses.

Tackling the challenge from both ends

Getting to a natural language dialogue state with a chatbot remains a challenge and will require a number of research breakthroughs. At FAIR we have chosen to tackle the problem from both ends: general AI and reasoning by machines through communication as well as conducting research grounded in current dialog systems, using lessons learned from exposing actual chatbots to people. Our strength lies in embracing the diversity that spans both approaches. From long-term, fundamental research like the CommAI initiative, to shorter-term applied efforts such as FastText or Facebook M. Through these, combined with our team’s expertise across the AI spectrum, from deep learning for NLP to reinforcement learning, computer vision, and engineering, we hope to deliver significant natural language dialog advancements.

An important aspect of FAIR work on dialog is how we ground it in clear foundations:

• Strong baselines: advanced learning systems for NLP problems should effectively deliver a performance improvement relative to more traditional methods. To that end, we built FastText with the goal of providing the very best results achievable with relatively simple and well understood techniques.
• Clear(er) evaluations: evaluating dialog systems is a hard problem. At FAIR we came up with better tools to do just that. At ICLR 2017, we are sharing findings and tools with the academic community on dialogue of our initiatives. These include the CommAI environment [1] to train and evaluate reasoning models, and the bAbI tasks for dialog [4] which can be used to test end-to-end dialog models. Thanks to our collaboration with Facebook M, these tools have been tested with models in real production conditions.
• Open research: FAIR publishes almost all its research in conferences or through preprints. Similarly, code and data, including the two evaluation initiatives cited above are released as open-source. Just as there is diversity of work in FAIR, there is also tremendous diversity across the AI community. We believe that open dialog and shared tools and learnings will lead to bigger advancements overall.

Making progress through shared learnings

At ICLR, we are presenting 7 papers that illustrate the quality, innovation and breadth of FAIR dialog research. Lazaridou et al. [6] and the CommAI team [1] propose directions for having systems be able to discover and use basic communication skills, a first step towards artificial general intelligence. Li et al. present two papers that study how end-to-end dialog systems can be improved by using ongoing live conversations to improve themselves [2, 5]. Bordes et al. introduces the bAbI tasks for dialog [4] to test end-to-end dialog systems in goal-oriented scenarios. And we present two papers on machine reading by Grave et al. [3] and Henaff et al. [7] that push the boundaries of text understanding by machines.

References
[1] CommAI: Evaluating the First Steps Towards a Useful General AI, M Baroni, A Joulin, A Jabri, G Kruszewski, A Lazaridou, K Simonic, T Mikolov
[2] Dialogue Learning With Human-In-The-Loop, J Li, AH Miller, S Chopra, MA Ranzato, J Weston
[3] Improving Neural Language Models with a Continuous Cache, E Grave, A Joulin, N Usunier
[4] Learning End-to-end Goal-oriented Dialog, A Bordes, YL Boureau, J Weston
[5] Learning Through Dialogue Interactions, J Li, AH Miller, S Chopra, MA Ranzato, J Weston
[6] Multi-Agent Cooperation and the Emergence of (Natural) Language, A Lazaridou, A Peysakhovich, M Baroni
[7] “Tracking the World State with Recurrent Entity Networks,” M Henaff, J Weston, A Szlam, A Bordes, Y LeCun