Over the last two years, machines have demonstrated their ability to read, listen, and understand English well enough to beat the best at Jeopardy!, answer questions via iPhone, and earn college credit on college advanced placement exams. Today, Google, Microsoft and others are rushing to respond to IBM and Apple with ever more competent artificially intelligent systems that answer questions and support decisions.
What do such developments suggest for the future of education?
Even excluding political dynamics, these are turbulent times for education, including for academic institutions, textbook publishers, and educational technology providers. Ubiquitous computing on increasingly capable phone, tablet, and cloud-computing infrastructures in an increasingly open environment are driving courses and textbooks on-line and onto hand-held devices. As if these disruptions are not overwhelming in themselves, the incorporation of artificial intelligence within emergent educational technology promises further disruption albeit with innovation that actually improves learning beyond “merely” democratizing education.
What is the business case for self-understanding textbooks?
Consider a machine that “understands” its content. In Project Halo, we found that educating a machine such that it could get college credit in the sciences cost on the order of $10,000 per age a decade ago. Over time, using graphical metaphors, the cost per page was reduced to a few thousand dollars. And, more recently, by understanding English using the Linguist™, the cost has fallen by two orders of magnitude.
A page typically contains less than 25 sentences of well under 20 words in length each, on average. Using statistical natural language processing, these sentences are too complex to be precisely understood, although evidentiary approaches, such as demonstrated in Watson, suggest that most can be understood with sufficient precision to answer many questions. As the questions become more difficult, however, such as on AP exams or whenever several steps of inference need to be “chained” together, it becomes necessary to understand more sentences more precisely. Today, understanding a sentence accurately requires a human being to resolve various ambiguities that induction alone cannot.
To disambiguate a typical textbook sentence now requires somewhere between seconds and minutes. Doing so results in understanding more of the subject matter more deeply and more accurately than Watson can. To do this for a page requires an hour or so. The result is artificial intelligence that can answer questions that are far more difficult than Watson can answer and doing so reliably rather than guessing. Moreover, the reasoning technology involved can explain the chain of inference and relate it to the words, phrases, and sentences from which the knowledge was acquired. (In today’s electronic publishing platforms and delivery vehicles, such indexing and richer navigation is done manually.) And the skill required to help the machine understand subject matter is widely available. People who contribute to Wikipedia are able to clarify the meaning of their text using the Linguist on day 1.
Consider another perspective. Question banks are commonly used in education and in test preparation. Question bank contributors can now add the sentences of knowledge needed to prove or refute answers. By assisting the machine in understanding those sentences, the machine is able not only to answer such questions on its own but to assess and assist learners in understanding the subject matter such that they do better, too.
This is actually easier than it may seem, as has been demonstrated by a prototypical electronic book that lifted the test scores of college biology students. In that case, the machine simply presented suggested questions that students could answer and through which students could navigate through relevant textbook content. In effect, the textbook integrated a question bank for which the machine had pre-computed the proofs and refutations of correct or incorrect answers. The logical semantics of these proofs was linked to the textbook content and as the basis for automatically rendered conceptual summaries.
What are the implications for educational technology markets?
Many textbooks generate millions of dollars of revenue per annum. And yet these textbooks are vulnerable to increasingly open publishing of electronic text. Although electronic textbooks offer more value, convenience, and functionality than the paper can offer, they are generally passive with regard to learning. Electronic textbook delivery platforms that incorporate active, adaptive learning such as enabled by machine understood content and artificially intelligent question answering, assessment, and pedagogical capabilities offer significantly more value to learners that either form of textbook alone. Now, a relatively small incremental investment can secure competitive distinction that avoids commoditization by electronic publishing. Thus, all publishers, including copyright holders and delivery channels are motivated to adopt understanding.
But what are the implications for tomorrow? Will textbooks be written in a more knowledge-driven manner rather than as sequential discourse? Perhaps. But perhaps that’s the wrong question. What clearly matters is whether learners learn better. There can be little doubt that content that understands itself and artificial intelligence that answers and explains questions better than Watson will engage and teach learners better.