Most teachers and educators consider reading comprehension to be applicable to human students, usually starting in about the first grade. However, there is a fast-growing type of comprehension that has nothing to do with humans. It’s called “Machine Reading Comprehension (MRC),” and it reflects an artificial intelligence (AI’s) ability to understand specific knowledge found in a wide variety of sources. This is a skill required of machines for many real-world scenarios. One such example would be for search applications, where it is preferable for an AI to give a precise answer rather than to simply provide a URL for the webpage that contains the sought-for information. Another, still not yet possible, example would be for the AI to help a doctor find information among thousands of documents; this decreases the amount of time the doctor would have to spend on such a task, and it has the potential to improve treatment speed and outcomes. In both cases, the AI would have to understand the nature of the document’s contents in a very specific way.
Current MRCs, though, are created based on supervised training data, which means they are trained using not only the articles they are supposed to understand, but also human-inputted questions about those articles with the corresponding answers. This process has severe limitations, because the labeling process must be done for each and every domain of knowledge. For example, in order to help doctors with a document search, it would be necessary to create one MRC for each disease, and each MRC would have to be constantly updated to reflect the ever-increasing number of articles being produced on that specific topic.
Microsoft is now researching a new “two stage synthesis network” models for training MRCs, called SynNet. SynNet is first taught key knowledge points, or semantic concepts, from a single domain based on the supervised data available. Then, it learns to form its own natural language questions (English, not code) around these potential answers within the framework of a given article. Once trained, SynNet can be used in an entirely new domain to generate dummy questions and answers for a given article – that enables it to create the supervised training data required for training specific MRCs usually provided by human input. By removing the human component SynNet becomes, for all intents and purposes, an AI teacher of document comprehension.[1]
That teaching is like a human teacher, who, based on her experience in previous domains, creates questions and answers from articles in the new domain, and then uses these materials to teach her students to perform reading comprehension in whatever the new domain may be. On the flip side, Microsoft has also developed a set of neural machine reading models, such as ReasoNet, which are akin to the students who learn from the teaching materials to answer questions based on the article with which they’re presented.
SynNet and ReasoNet are still in the early stages of development, and it remains to be seen whether full reading comprehension, a necessary skill for AIs to achieve the utmost goal of general intelligence, is actually possible. Though packed with possibilities, not everyone is happy with the learning, comprehending MRCs. Some envision them turning into a version of Skynet, the intelligent machine of the “Terminator” movies that tried to wipe out humanity. Businessman and space pioneer Elon Musk has warned U.S. Governors that they need to regulate such AIs “before it’s too late.” And theoretical physicist Stephen Hawking has warned that intelligent AIs may lead to the downfall of our species.
Regardless of which side you’re on, the fact is we have a rudimentary system in place in which a machine can learn to understand the words it reads. Who knows what progress that will help human teachers make when instructing their own students in comprehending reading material.
Do you have an opinion on this subject? Please leave us your thoughts in the comments section below.
Citation:
[1] Microsoft Blog. (June 26, 2017). Transfer learning for machine reading comprehension. Retrieved from https://www.microsoft.com/en-us/research/blog/transfer-learning-machine-reading-comprehension/
