Behind Microsoft's Nuance Deal: Natural Language Processing, Explained
By Jared Council
Microsoft Corp.'s $16 billion acquisition of Nuance
Communications Inc. will bolster its strength in natural language
processing, artificial intelligence that can interpret and respond
to human language.
The technology has become ubiquitous in recent years. It powers
Apple Inc.'s Siri and Amazon.com Inc.'s Alexa and automated
call-center attendants. Voice is coming to rival text as a
computing interface. Research firm Gartner predicts that by 2025,
50% of knowledge workers will use an NLP-powered virtual assistant
daily, up from 5% in 2020.
Here's how natural language processing works.
Calculating the odds
Much of today's natural language processing calculates the
probability of how a sequence of words should be interpreted among
all the possible ways. For instance, it can tell whether the word
"club" likely refers to a sandwich, the game of golf, or night
When and where did it originate?
The field traces its roots back to the 1950s and 1960s. Its
pioneers include linguist Noam Chomsky and computer scientist John
McCarthy, according to Michael Picheny, a research professor at New
York University's Courant Institute Computer Science Department and
Center for Data Science. In its early days, the process of getting
computers to analyze and understand language involved hard-coded
software rules and expert systems, according to Dr. Picheny, who's
been working in the field since 1979.
Enter neural networks
The technology has advanced with the maturity of so-called
neural networks, which began to take off in the early 2010s,
according to Dr. Picheny. Neural nets mimic neurons in the human
brain, and power an AI technique known as deep learning, as well as
Whether it's trying to understand text or speech, a
deep-learning model analyzes a sentence by first examining
fundamental components known as tokens, which can be words or
segments of words, Dr. Picheny said. Those tokens are analyzed in
tandem with other words in the sentence to determine the likely
usage of each word. A neural network trained on millions or
billions of sentences can make a probabilistic assumption about
which kind of bank is being talked about in the sentences "I went
across the river to the bank" and "I went across the street to the
bank, " for instance.
If a speaker is trying to book a flight with an NLP-powered
automated system, the algorithm looks for words that are likely the
origin city, the destination city as well as the departure and
return times, Dr. Picheny said.
The use of NLP in voice assistants tends to involve additional
steps, according to Bern Elliot, a research analyst at Gartner Inc.
The system identifies the words based on the sounds and makes
statistical guesses about what word was uttered based on
surrounding words. Those words are translated to text and analyzed
for intent, such as whether the word "I" or "eye" has been uttered.
In communicating a response to the user, the text is then converted
back to speech.
What are the challenges?
NLP systems currently in use are designed for very specific
uses, and their performance tends to break down when they encounter
something outside of that zone, said Ian Jacobs, an analyst at
Forrester Research Inc. For example, if a chat- or voice-bot is
helping a user make a flight reservation, that system may have
trouble answering questions about reward points or the status of a
It is also difficult to train NLP systems to predict which
situations might require a simulation of empathy and convey that
response in the choice of words and tone, which can affect the user
experience, Mr. Jacobs said. "You don't want to say something to
the system like, 'Yeah, my mom just died,' and then have the system
say, 'Oh, great. Your mom just died.' It's got to mirror your tone
as well," Mr. Jacobs said.
What does the future hold?
Academic and corporate researchers are striving to make NLP
systems more general -- that is, able to handle more requests and
inquiries without breaking down. Another goal is to make the
systems more responsive to context. For instance, Amazon said in
December that its Alexa technology can ask a clarifying question to
understand requests it has never heard before. Right now, that
capability is limited to a few applications, but they are likely to
expand. Earlier this year, software company OpenAI LP showed that
its NLP system, dubbed DALL-E, could generate images based on a
user's request -- for instance, it could produce an image of "an
armchair in the shape of an avocado" if instructed to.
Researchers see NLP as one of the most promising technologies
for creating machines that can exhibit what is known as general
intelligence by performing tasks and responding to a broad range of
human requests without being explicitly trained to do so.
"Though scientists and researchers have done a lot of
theoretical work on NLP in the past, we have only recently started
seeing its real-world use cases," said Ritu Jyoti, who leads the AI
practice at International Data Corp.
Write to Jared Council at email@example.com
(END) Dow Jones Newswires
April 13, 2021 13:17 ET (17:17 GMT)
Copyright (c) 2021 Dow Jones & Company, Inc.