We are using the developed English speaking ability assessment system to develop a framework for providing both students and educators satisfactory feedback by combining knowledge gained from analysis of large-scale real-time English conversation data with supervisory guidance from English language education experts to elucidate and explain linguistically how the English language abilities of students participating in online English conversation classes develop in the short and long term.

 

Long-Term Vision: Conversation AI that Evolves with People

Artificial intelligence (AI) has long been discussed as a threat to human professions and/or human rights. However, AI is strongly dependent upon the ethical values of its creators. How should AI be designed to coexist in harmony with people and expand the capabilities of society as a whole in the future? Here, we assume the following are some of the requirements for the social AI to be accepted by particular groups or communities:

1. Compliance with social norms: Ability to comply with social norms (protocols) [1].
2. Earning Trust: Being able to win trust as a social participant by being accountable and continuing to behave predictably and consistently [2] [3] [4].
3. Achievement of agreed objectives: Being able to achieve objectives through cooperation with others while complying with the social norms listed above, reaching agreement democratically, and recognizing the common values that are called for in particular situations.

In order to make a solid step toward realization in the case of either narrow or wide AI, it is necessary to find a good domain and an appropriate evaluation function. For example, the success of Google DeepMind’s AlphaGo in the past has been attributed to the fact that they chose a domain with a clear value of Go [5]. What about the social requirements for conversation AI listed above? Matsuyama, the principal researcher, has long been researching robots that can facilitate conversations with large numbers of people and conversation agents that can build relationships with others. However, these are still limited in terms of individual and basic ability to realize social integrity. The fields of speech language processing and dialog systems research have long and detailed histories. The emergence of deep learning has led to the further development of the underlying technologies for speech recognition, language comprehension, and generation. However, there has been no fundamental progress on methods to incorporate complex conversational AI components.

Today, the concept of the CEFR in English-speaking education can be said to be a broad and comprehensive evaluation criterion for how the quality of communicative skills and strategies should be evaluated in specific tasks when we humans are viewed as social agents. For example, the task of interviewing is a communicative task that involves an adequate understanding of another’s identity. Understanding the identity of others requires a variety of basic communication skills and advanced strategies. Presentations are tasks that make persuasive arguments based on objective data. Negotiations are tasks where a common goal is reached through understanding and agreement. In this regard, we believe that the CEFR is a framework that is suitable as both a foundation for language learning and a very reasonable design direction for conversational AI.

The Uniqueness of the Research

The uniqueness of the current study is its proposal of a design paradigm for a social AI which grows with students and tutors when the English language conversation class is considered a small society. As mentioned earlier, the difficulty of conversational AI design is generally due to the difficulty of collecting large amounts of good conversational data and the complex and arbitrary nature of performance evaluation. However, in this study, we have chosen English conversation classes as a domain for which large-scale data collection is easy. Moreover, since it is anticipated to be reliable, high-quality teacher data can be collected from already established performance evaluation indicators of the CEFR. Advantages are guaranteed when compared to conventional conversational AI and educational engineering research. In addition, the CEFR includes the three requirements of social AI described above (compliance with social norms, obtaining trust, and achieving goals based on agreement). Research applying these to the design of social AI which can coexist with humans is unprecedented worldwide.

Outlook in A Larger Perspective

A conversational AI which could, for example, follow and facilitate a one-hour conversation task (equivalent to a class in Tutorial English) without breaking down can be developed by limiting it to a particular communicative ability or particular situations if there is an ingenious machine learning algorithm and sufficient data. As such, we believe we will approach the realization of a generic social conversational AI that can coexist with the human society by viewing it as something which is positioned equally with humans within the social environment, and by completing important milestones one by one while comprehensively evaluating each language communication capability and strategy for each particular task. We can expect the following practical, real-world benefits.

• Benefits for educators: An automated proficiency assessment system helps tutors perform simultaneous, equitable, and rapid progress evaluations of multiple students. This means a reduced cognitive load on the tutor in class, resulting in the provision of high-quality, inexpensive tutoring services to our students.
• Benefits for learners: While improving language skills requires both training quality and quantity, with the advent of a fully automated AI Tutor, students would be able to take as many high-quality, inexpensive English conversation classes as they want, whenever they choose. It will likely be of benefit to beginners developing basic skills.
• Benefits for the whole of Japanese society: The AI English conversation assistance service which will be realized as a result of the current research will directly enhance the English communication skills of the Japanese people and foster the ability to live in an increasingly diverse world. Consequently, Japan’s presence on the world stage will be greatly enhanced, benefiting the national interest.

As outlined above, this research topic is expected to contribute to the realization of a future society in which the benefits of AI technology development can be maximized by everyone in society through basic research for realizing social AI with practical English conversation as the domain.

1. Tetsunori Kobayashi and Shinya Fujie. Conversational robots: An approach to conversation protocol issues that utilizes the paralinguistic information available in a robot-human setting. Acoustical Science and Technology, 34(2):64–72, 2013.
2. Niklas Luhmann, Trust and Power, 1982.
3. 山岸俊男, 小宮山尚, 信頼の意味と構造 — 信頼とコミットメント関係に関する理論的・実証的研究, INSS Journal. 2 pp. 1-59, 1995.
4. Timothy Bickmore and Justine Cassell, Relational agents: a model and implementation of building user trust, In Proceedings of the SIGCHI conference on Human factors in computing systems, pp. 396-403. ACM, 2001.
5. Fei-Yue Wang, Jun Jason Zhang, Xinhu Zheng, Xiao Wang, Yong Yuan, Xiaoxiao Dai, Jie Zhang, and Liuqing Yang, Where does AlphaGo go: From church-turing thesis to AlphaGo thesis and beyond, IEEE/CAA Journal of Automatica Sinica 3, no. 2, pp. 113-120, 2016.