Inquisitive artificial intelligence that asks questions about things it reads could be used to quiz students in class. The question-asking ability would also help chatbots with the back and forth of human conversation.

AI is usually on the receiving end of queries, says at Cornell University in Ithaca, New York. Du and his colleagues have turned the tables by building a system that has learned to ask questions of its own.

This is something that people have been wanting to do for a long time, says at the University of Dallas in Texas. Previous attempts by other people using hand-coded rules haven’t been particularly successful.

The machine-learning algorithm can read a passage of text and come up with the kind of questions you might ask to check someone’s understanding of a topic. Du’s team used a neural network – software that loosely mimics the structure of the brain – and trained it on more than 500 Wikipedia articles and 100,000 questions about those articles sourced from crowdworkers. For example, a sentence about different types of crop grown in Africa might be paired with the question “What is grown in the fertile highlands?”

Pattern recognition

The software learned to recognise patterns that linked questions back to their source text, such as that dates in sentences tended to correspond with questions that started with “when”, and sentences about places often led to questions that started with “where”.

The team then presented the AI with extracts from Wikipedia articles that it hadn’t yet seen. Given a passage from an article on economic inequality, it asked the question “When did income inequality fall in the US?” An article about a Pakistani political organisation prompted “When was the Jamaat-e-Islami party founded?” Four human volunteers rated the naturalness and difficulty of the questions as higher than those generated by existing systems.

Mazidi is impressed with and thinks similar algorithms could eventually be used in classrooms to help test students. “If you stop a student while they’re reading something, and ask them a few questions, it can greatly improve their comprehension of what they’ve just read,” she says.

Du’s software taught itself using a large amount of real-life data, but Mazidi thinks rule-based approaches shouldn’t be thrown away entirely. “A neural network is going to learn whatever it learns and you have very little control over that,” she says. By adding in a few rules, a computer could be made to formulate questions in a way that seems more human.

The current version of the system produces a question for every sentence it reads but Du wants it to asks questions only about sentences that contain statements. “Not all sentences are question-worthy,” he says.

Reference:

Read more: Google uses neural networks to translate without transcribing; Kindergarten bots teach language to tots