Alibaba AI to break the world record reading comprehension accuracy rate First super human

Click to have a surprise

At the beginning of the 2018, artificial intelligence made a major breakthrough. Squad, the top event in the field of machine-reading comprehension, launched by Stanford University on January 11, has cheered the industry on the first time that artificial intelligence has surpassed humans in its history of reading. Alibaba has broken the world record with a 82.440 accuracy rate and surpassed 82.304 of human achievements.

Squad in charge of Pranav Rajpurkar difficult to cover the excitement of feeling. He said in social media that 2018 was a strong start, and the first model (SLQA + submitted by Alibaba IDST team) surpassed human performance in matching accuracy. Next challenge: Fuzzy match, Human still lead 2.5 points.

The squad competition builds a large-scale machine reading comprehension dataset (with 100,000 questions), which comes from more than 500 articles from Wikipedia. After reading a short article in the dataset, Ai needs to answer a number of questions based on the content of the article, and then compare it with the standard answer to get the result of an exact match (Exact match) and a fuzzy match (F1-score). Squad is the industry's recognized top event in machine reading comprehension, attracting the deep involvement of well-known corporate research institutions and universities including Google, Mellon University, Stanford University, Microsoft Research Asia, Allen Institute, IBM, and Facebook. The breakthrough in this technology stems from the Alibaba research team's "layered fusion attention mechanism" of the deep neural network model. The model can simulate the behavior of human being in the reading comprehension problem, including examining the content of the text, reading the article with questions repeatedly, avoiding forgetting in the reading and making the related annotation. The model can capture the problem and the specific area of the article in the same time, with the help of layered strategy, focus gradually to make the answer boundary clear; On the other hand, in order to avoid too much attention to detail, the integration of the global information into the attention mechanism, to ensure that the focus is correct. Sro, chief scientist at Alibaba's natural language processing, said that the machines have achieved very good results in solving the wiki-type objective knowledge question, and we will continue to move towards the ultimate goal of "understanding and thinking" about common content. In the future, the focus of research and development is to apply this technology to the actual scene, so that the machine intelligent Pu hui life.

In fact, the technology has been widely used within Alibaba. For example, a large number of customers are consulted on the rules of activity every year, with a double 11. Ali Xiaomi team through the use of the technology of the Luo team, so that the machine directly to read the rules, to provide users with rules interpretation services, is the most natural way of interaction.
For example, customers will also ask a lot of basic questions about a single item, and these questions are actually answered on the Product Details page. Now through the machine reading comprehension technology, can let the machine to the detail page the product description text to read and answer more intelligently, reduce the service cost while increasing the purchase conversion rate.
The natural language processing team, led by Sro, supports the entire ecological technology needs of Alibaba. The ALINLP natural language technology platform, which was developed by them, calls 120 billion + times a day, and the Alitranx translation system offers 20 languages online service daily calls of more than 700 million + times. Prior to the 2016 ACM CIKM personalized Electronic Business search, 2017 ijcnlp Chinese grammar Test cged evaluation, 2017, the United States Standards and Metrology Bureau TAC Rating English entity classification and other competitions to achieve the world's first results.

Click to have a surprise