AI tutors will make mass retraining a viable reality

Artificial intelligence may be threatening employment but it could also be key to helping humans find alternative jobs, argues Shigeru Miyagawa 

May 22, 2019
Source: iStock

It is commonly agreed that automation will take over large numbers of existing jobs over the next generation, requiring humans to train and retrain for new but different roles.

We can already see this happening. College graduation rates are increasing across all Organisation for Economic Cooperation and Development (OECD) countries, and one-third of this increase is accounted for by adults aged between 25 and 34, who are already likely to be working. Many will be taking online courses; even in a country such as Japan, in which the college-age population is declining, enrolment is expanding on the few fully online programmes available.

There is no shortage of high-quality higher education material available online: this movement began in 2001 with MIT’s OpenCourseWare, which today has material from 2,400 MIT courses and is accessed monthly by 2 million unique users. And such material is drawn on by the burgeoning number of massive open online courses, which allow students to learn just about anything. According to Mooc search engine Class Central, last year saw 101 million learners worldwide studying 11,400 Moocs offered by 900 universities.

There are AI programmes that can guide you through the myriad offerings fro some topics and find the appropriate set for your needs. But what about once you enrol? One well-known problem with Moocs is their high attrition rate, especially in the first weeks. To offer the best online learning experience, high-quality content clearly isn’t enough.

Learners tend to only passively engage with online content, while effective education requires active learning. This is borne out in a recent study by an MIT undergraduate, who found that those who stayed the course on a Mooc on modern Japanese history became increasingly active on the course’s online discussion forum.

Armed with this knowledge, I approached a colleague in MIT’s AI Lab, Boris Katz, to collaborate on a 24/7 AI tutor for Moocs using the system for answering natural language questions that he had already developed. The software, called START, converts the question into a semantic representation (common for all wording variations), which is used to search the web or whatever other source we point it towards, to come up with, ideally, one correct answer.

Our team annotated the video lectures such that if the answer to their question can be found in one of them, the student is taken directly to it. The answers to more general questions are retrieved via general web sources, such as Wikipedia infoboxes.

The effort involved in creating even this simple prototype was enormous. But as we learn to automate more of the tasks, such as annotation, the effort will diminish. And it will pay off: automatically dealing with Moocs’ basic knowledge content will allow interhuman discussion to focus on higher-level issues.

It is also important for the future workforce to acquire effective communication skills. One way is to get students to write clear, logical essays, but these are difficult to mark at scale. One solution that has been used since the early Moocs is peer grading, but it has met with mixed results. One informal test found that peers tend to be overly generous, mostly giving As. However, an AI marker assigned the same small set of essays a range of grades with a similar distribution to those given by human teaching assistants.

The problem is that setting up an AI essay grading system effectively is also quite complex because it needs to be fed all the course reading materials and lecture notes to ensure that it is grading in the context of the course’s full knowledge base. The instructor’s rubrics must also be fed in and weighted accordingly.

However, again, greater automation could make AI essay grading a reality, bringing online learning one step closer to meeting the needs of learners in ways that are only currently possible in face-to-face contexts.

We still have some way to go before we can implement AI systems in real time. But, ultimately, the goal of large-scale, affordable, lifelong learning will become a reality.

Shigeru Miyagawa is senior associate dean for open learning at the Massachusetts Institute of Technology. He is a keynote speaker at THE’s Teaching Excellence Summit, to be held from 4 to 6 June at Western University, Canada.

后记

Print headline: Tech unburdens teachers

登录 或者 注册 以便阅读全文。

请先注册再进行下一步

获得一个月的无限制地在线阅读网站内容。只需注册并完成您的职业简介.

注册是免费的,而且非常简单。一旦成功注册,您可以每个月免费阅读3篇文章。:

  • 获得编辑推荐文章
  • 率先获得泰晤士高等教育世界大学排名相关的新闻
  • 获得职位推荐、筛选工作和保存工作搜索结果
  • 参与读者讨论和公布评论
注册

相关文章

欢迎反馈

Log in or register to post comments

评论最多

Recent controversy over the future directions of both Stanford and Melbourne university presses have raised questions about the role of in-house publishing arms in a world of commercialisation, impact agendas, alternative facts – and ever-diminishing monograph sales. Anna McKie reports

3 October