SQuAD's head, Pranav Rajpurkar, is struggling to hide the excitement, saying on social media that 2018 was a strong start. The first model (SLQA submitted by the Alibaba iDST team) surpassed human performance in terms of accuracy matching! Next challenge: fuzzy match, humans are still 2.5 points ahead!
SQuAD game to build a large-scale machine reading data set (including 100 thousand), articles from the more than 500 Wikipedia article.
After reading a short paper on data set, AI needs to answer several questions based on the content of the article, then compare it with the standard answers, and get the results of exact matching (Exact Match) and fuzzy matching (F1-score).
SQuAD is the industry's recognized machine reading comprehension top competition, attracting Google, Carnegie. Mellon University, Stanford University,MicrosoftThe Asian Institute, the Alan Institute, the IBM, Facebook and other well-known enterprise research institutions and universities are deeply involved.
The major breakthrough in this technology comes from the deep neural network model based on the hierarchical integration of attention mechanism proposed by the Alibaba research team. The model can simulate human behavior in reading comprehension problems, including the combination of text topic, repeated reading with questions, avoid reading forgotten and related notes.
The model can capture problems in specific areas of association and articles at the same time, with the help of hierarchical strategy, gradually focus, make the answer clear boundary; on the other hand, in order to avoid too much attention to details, the fusion will join the mechanism of attention to global information, appropriate correction, to ensure the correct focus.
The chief scientist of the Auror Alibaba Natural Language Processing said, to solve the question of Wiki objective knowledge, machines have achieved very good results, we will continue to forward the ultimate goal of universal content "to understand thinking".
In the future, the focus of R & D is to apply this technology to the actual scene, and make the machine intelligent life.
In fact,This technology has been widely used within the Alibaba. For example, a large number of customers have a large number of customers every year to consult the rules of the activity. Ali team by using the Auror team Xiaomi technology, make the machine directly on the rules of reading, reading provides regular service for users, is the most natural way of interaction.
For example, customers will also ask a large number of basic questions for a single commodity, which are actually answers to the details of the product. Now, through machine reading and understanding technology, the machine can read and answer the texts more intelligently, and reduce the cost of services and improve the purchase conversion rate.
The Natural language processing team, led by Siro, supports Alibaba's entire ecological technological needs. Their AliNLP Natural language Technology platform is invoked 120 billion times a day. The Alitranx translation system provides over 700 million daily calls in 20 languages.
Previously, it won the world's first achievement in the 2016 ACM CIKM personalized business search, 2017 IJCNLP Chinese grammar test CGED evaluation, and the 2017 annual US Standard Bureau TAC comparison of English entity classification and other competitions.