It’s no secret that Microsoft is very focused on applying the set of technologies loosely defined as “AI” to its products and services. The company’s combined AI + Research unit, which is now up to 7,500 employees, is evidence of that.
But it’s still interesting to see the types of projects the company’s researchers deem ready to show off at industry conferences and internal confabs, as this often provides clues as to what Microsoft may be ready to try to commercialize next.
One of these areas where the company is upping its focus is in machine reading, or the automatic understanding of text. At Microsoft’s Faculty Research Summit in Redmond this week, officials shared a glimpse of what Microsoft’s doing on this front.
On July 17, two Microsoft researchers presented on machine learning. One of these researchers, Jianfeng Gao, Partner Research Manager, also is one of the authors on a paper on a new neural network architecture in which Microsoft has been investing, called ReasoNet.
ReasoNet, short for the Reasoning Network, is targeted at machine comprehension. “ReasoNets make use of multiple turns to effectively exploit and then reason over the relation among queries, documents and answers,” according to an abstract about the paper, which the researchers will present at the August SIGKDD Conference on Knowledge Discovery and Data Mining.
The ReasoNet model is meant to mimic the inference process of human readers. Microsoft has applied the shared memory component of the model to Knowledge Graph Completion Task, according to the company’s Research Web site. The team also has developed a two-stage technique for transferring learning in machine comprehension that it calls SynNet — the subject of another research paper due to be presented at a conference in September this year.
ReasoNet is a project from Microsoft Research’s Deep Learning Group in Redmond. The Deep Learning for Machine Comprehension project, which Microsoft established in September 2016, has set its sights on teaching computers to read and answer general questions pertaining to a document.
Microsoft’s January 2017 acquisition of deep-learning startup Maluuba also plays heavily into what the company is doing on the machine-reading front.
Maluuba has been working on enabling machines to “comprehend, reason and communicate with humans,” as its website notes. The startup has pioneered ways to train machines to seek information and read and reason. The company also is finding ways to train machines to ask questions.
Maluuba’s machine-reading comprehension (MRC) system can ingest a 400-page auto manual and then answer user questions based on it in real time.
“The long-term vision for this product is to apply MRC technology to all types of user manuals, such as cars, home appliances and more,” according to a Channel 9 video of the demo that’s part of the Research Faculty Summit collateral.
Last year, Microsoft released a free database of 10,000 questions and answers, Microsoft MARCO (for Machine Reading Comprehension), as part of its machine-reading efforts. These kinds of massive datasets are essential for training machines.
Looking ahead, Microsoft seems to be counting on machine reading as a way to perfect the concept of having an expert assistant built into its products.
Microsoft already has Cortana built into products like Windows, but there have been persistent rumors that the company is looking to go further with something like a personal business assistant — perhaps in the form of a Bing Assistant or concierge bot.
In an interview with GeekWire earlier this year, Microsoft founder Bill Gates mentioned the importance of reading text to the company’s future products and services.
“When you go to look at communications you shouldn’t have to just look at a timed order fashion, you should trust that it’s understanding of you and the context and priorities are there. But only by reading that text will we do that, so there’s a frontier here that’s very exciting that Rajesh Jha (the head of Office applications), Harry Shum (the head of Microsoft’s AI + Research group), a lot of the key people under Satya are grabbing onto that, and some particular opportunities around that are where the resources are being shifted,” Gates said.
The biggest hurdle for machine reading, according to the researchers presenting this week at the Faculty Research Summit, is introducing the element of common sense into machine-reading situations. Humans, being multi-modal, have a variety of ways to filter and understand things that are inherent in written text, the researchers noted.
While that work continues, the next chapter in machine reading at Microsoft is being written… and read.