Chinese institute wins machine reading competition

The Joint Laboratory of Harbin Institute of Technology and iFlytek (HFL) has won the top prize at the Stanford Question Answering Dataset (SQuAD) Competition with its newly-developed system, reports the thepaper.cn.

Photo shows a product release conference of iFlytek [Photo: Baidu]

HFL completed its machine reading program in May, 2015. It's said to be able to read at the level of a 6-year old.

"While it is easy for machines to memorize and store mass information nowadays, the major challenge is to make them accurately interpret the text," said Wang Shijin, vice director of HFL.

Known as the "ImageNet of machine reading comprehension", the SQuAD Competition attracts many of the world's top academic and commercial institutes, such as IBM, Facebook, Google, CMU, Stanford and more. Microsoft Research Asia was the winner last year.

In this year's competition, every system model was given a short article, then had to answer five questions. HFL's model scored in a C+ to B+ range.

HFL's model is equipped with a new type of neural network called "attention-over-attention (AoA)," which is intended to eliminate the possible errors resulting from natural language processing (NLP), said Wang Shijin.

HFL system is being tested to see if it can accurately mark College Entrance Examinations, in the hope of one day relieving teachers of the often-cumbersome requirements involved with grading papers.