Machine Outperforms Humans On A Reading Test
A deep neural network model developed by the Institute of Data Science of Technologies, the research arm of Alibaba, outdid human participants in the Stanford Question Answering Dataset or Squad.
Alibaba's AI learning model that can read from paragraphs to sentences to words was based on the Hierarchal Attention Network, making it very similar to the natural human language.
"This is the first time that a machine has outperformed humans on a such a test," said Si Lou, chief scientist of natural language at IDST. Human participants in the reading test scored 82.304 based on the January 5 tests.
A similar technology from Microsoft Research Asia scored 82.640, also beating the humans in Jan. 3 tests.
Alibaba's result was officially registered by Squad on Jan. 11, a day ahead of Microsoft's, technically making the Chinese company the first to beat human scores.
"These kinds of tests are certainly useful benchmarks for how far along the AI journey we may be," said Andrew Pickup, a spokesman for Microsoft. "However, the real benefit of AI is when it is used in harmony with humans."
Squad Reading Test
Stanford University AI experts developed the 100,000 questions test to measure the ability of machines to process a large amount of information and answer them with accuracy. The questions are generated based on hundreds of Wikipedia articles. That way, the AI model can only supply specific and precise answers.
Squad is regarded as the most comprehensive and authoritative machine-reading gauge. Tech companies such as Google, Facebook, Microsoft, IBM, and Samsung are also using it to test if the machine learning models they built can answer questions from the data set.
Squad developer Pranav Rajpurkar of the Stanford Machine Learning Program described Alibaba's feat in the reading test as "a strong start to 2018" in terms of AI development.
Using the said neural network that can accurately answer questions that offer clear-cut answers, humans intervention and input can be lessened in answering queries related to customer service. The learning model can identify questions asked by customers and look for the most relevant answers from a set of prepared materials.
Alibaba tested this AI technology on chatbot Ali Xiaomi used by retailers on the company's virtual stores, Taobao and Tmall. A similar technology that is also based on the network was used as well to answer voluminous inbound queries from customers during the November 2017 Singles Day sale in China. Alibaba raked in $25 billion in the said online shopping festival.
Natural language processing technology was the most dominant form of AI in consumer technology products with virtual assistant and voice activation features previewed during the CES 2018.