Does AI Really Speak Our Language?
Does
AI Really Speak Our Language?
Artificial
intelligence has come a long way in handling tasks that involve language, but
it's still missing most of the nuance and context that we humans take for
granted.
By
Ben Dickson December 17, 2018 8:00AM EST
Thanks to advances in deep learning—the
subset of AI that solves problems by finding correlations between data
points—computers have become very good at handling tasks that involve the
processing and generation of language.
Robots
that look and sound human, digital assistants that can place calls on our behalf and
book restaurant tables, email apps that can predict and complete our sentences
with decent accuracy, translator apps that can give you near-real-time
translations, bots that can generate news stories—they're all powered by
artificial intelligence.
All these
developments can give the impression that AI has mastered language, one of the
most complicated functionalities of the human mind. But there are fundamental
differences between the way humans and AI handle language, and not
understanding them can create false expectations about AI's language-processing
capabilities.
Advances in the Field Are Real
Earlier
approaches to natural language processing (NLP) involved a lot of manual
coding. For instance, if you wanted to build an application that translated one
language to another, you had to create an entire software library that defined
the rules to extract the function of different words in a sentence, map those
words to their target language, and generate a sentence that was both
semantically and grammatically correct. The process was long, painful, and
error-prone.
Deep
learning uses a fundamentally different approach to machine translation.
Translations based on deep learning and neural networks "extract a large
number of latent patterns that significantly improve the quality of sentence
translation over previous approaches," said David Patterson, head of AI
for Aiqudo, a company that develops voice-enabled digital assistants.
The
fundamental component of deep-learning algorithms is the neural network, a
software structure roughly fashioned after how the human brain acquires
knowledge. Give a neural network a set of English sentences and their
corresponding French translations, and it extracts the common patterns between
the examples and uses them to translate new sentences it hasn't seen before. So
there's no need to hard-code rules.
In 2016,
the New York Times ran
a story in which it
detailed how a switch to deep learning had revolutionized Google's translation
service. "The A.I. system had demonstrated overnight improvements roughly
equal to the total gains the old one had accrued over its entire
lifetime," the Times wrote.
"When large amounts of training data are available,
deep learning has done very well; for example, in machine translation, as
compared to earlier statistical machine-translation models, [deep learning]
improved the quality significantly, even though there is still a need for
further improvements," said Salim Roukos, CTO of Translation Technologies
at IBM.
According to Roukos, on a scale of 0 to 5, neural
networks have helped improve machine translation by 0.5 to 1 point.
Beyond
translation, deep learning has helped make great advances not only in translation
but also in several other language-related fields, such as text summarization,
question-answering, and natural-language generation.
Several
companies have used deep learning to develop applications that enable users to
accomplish tasks through conversational interfaces such as messenger apps. In recent
years, AI-powered chatbots for banking, healthcare, customer service, and many other fields have
become very popular. Companies including Facebook and Google have created
platforms on which users can create smart chatbots without any coding—simply by
providing them sample phrases.
Deep
learning is also used in text-to-speech, the technology that enables us to send
voice commands to our computers, and voice synthesis, the technology that
enables computers to talk back to us in human-sounding voices.
There are
now several AI-powered voice-synthesis platforms that let users generate
natural-sounding voices. The technology can help solve different problems
ranging from company
branding to helping
ALS patients regain their voice.
Deep Learning's Limited Grasp of Language
"It
is possible to translate a sentence without understanding its meaning,"
Patterson said. This might sound absurd to a human translator, but for
computers, the difference is very real.
Machine translation has nothing to do with understanding
the source and target language. At its core, deep learning is a
pattern-matching algorithm. When it finds similarities between a new sentence
and the examples it has been trained on, it can provide an accurate
translation. Otherwise, it could produce meaningless results, as Patterson
pointed out.
"A
related issue is that of robustness of the system when there is drift (or
change) between the training data and the actual run-time data," Roukos
said.
Despite advances, all current blends of artificial
intelligence—including deep learning—are narrow
AI, which means they can
perform the very specific tasks they're trained on. Unlike humans, who can
generalize knowledge and transfer concepts from one domain to another, neural
networks perform poorly when applied to tasks that deviate from their
specialized domain.
This means
that while deep learning performs exceptionally well on simple translation and
language-related tasks, it can fail spectacularly when those tasks require
common sense and the understanding of abstract concepts.
Cognitive
science professor Douglas Hofstadter unpacked some of the limits of deep
learning–based translation in a feature published in The Atlantic in January. Hofstadter showed through
examples how an AI translation system would miss some of the simplest concepts
that a human translator would take for granted. For instance, any human reader
would deduce that the following excerpt describes the house of a married
couple: "In their house, everything comes in pairs. There's his car
and her car, his towels and her towels, and his library and hers."
Accordingly,
when translating the above excerpt to French, where possessive pronouns inherit
their gender and number from the possessed item rather than the possessor, a
human translator would add words and details to make sure that the reader would
understand that the text is describing a man and a woman. But in Hofstadter's
test, Google Translate generated a translation that completely missed the
point, omitting details that would highlight the
difference in gender between the two people who the original text was
describing.
"We
humans know all sorts of things about couples, houses, personal possessions,
pride, rivalry, jealousy, privacy, and many other intangibles that lead to such
quirks as a married couple having towels embroidered 'his' and 'hers.' Google
Translate isn't familiar with such situations. Google Translate isn't familiar
with situations, period," Hofstadter concluded.
"The
subtleties of meaning in context, especially with a longer text, is still a
challenging problem for machines. Also, there is a need to have more 'world
knowledge' to help systems understand language in a way that is similar to human
level," IBM's Roukos said.
Earlier this year, separate reports from The Wall Street Journal and Wired detailed
how several companies that used deep learning for services such as smart email
replies and intelligent chatbots were employing human operators to jump in and
take over when the AI failed—which happened quite often. Facebook also used
this practice with its abandoned M
chatbot assistant, and
Google is using it in the trial of its Duplex
technology, which makes
phone calls on behalf of users and books restaurants.
"We
can't hope that real consciousness will become part of AI in the near future.
For now, we need to use models of different real-world domains to help organize
human-computer conversation on the level that 'simulates' human-human
conversation," said Patterson. "In order to move from 'simulated'
conversations to 'real' conversations, we need to figure out how to empower
machines with things like common sense, understanding of context, and
creativity. This is where the current research is focusing."
Scientists
and researchers are working on different techniques that might complement deep
learning or enable neural networks to do more than just match patterns. The
hard truth is that natural language processing is a very complicated task. Some
believe true NLP requires general artificial intelligence, AI that is on par
with the human mind. And that is nowhere in sight. Until then, we have to
acknowledge the limits of deep learning to make the most out of the tremendous
progress we've made into teaching our computers to speak in our language.
Comments
Post a Comment