How to Train Your Computer
It’s how we share information with co-workers; teach our children how to ride a bike and write a thesis arguing the societal benefit of space exploration.
Linguists refer to our everyday language as "natural language" because it evolved naturally rather than by design. Deliberately created languages, such as those used by computers, are called "artificial languages."
These machine languages enable our computers to process structured data and information; names, addresses, numbers and the like. But imagine the potential of a computer that understands natural language. Armed with those language skills—and an understanding of how we use that language to answer questions—a computer could gain access to all the knowledge stored in our natural language. That goes beyond databases and traditional information sources. Now we’re talking about the unstructured data of emails, social posts, audio and more, things that computers often can’t process, but contain a wealth of valuable information.
That's natural language processing. It's the idea behind machines that can both understand and communicate in the way we speak. It's how Siri® knows to give us directions to Grandma’s house and Amazon's Alexa can play your favorite song. It's how IBM Watson beat humans in Jeopardy and now it's being used in significant, potentially life-altering endeavors such as diagnosing illnesses and recommending treatments.
Imagine the potential of a computer that understands natural language.
At State Street, we’re training machines to understand our clients’ questions, posed in natural language, of course, and respond with meaningful answers. Our goal is to help portfolio managers, chief investment officers, analysts and others wade through a complex field of financial information.
What’s driving inflation in the emerging markets? What are the risks to the euro over the next three months? How will Brexit impact investor confidence globally? Answers to these questions exist, but it’s difficult and time-consuming to find them.
Computers that can understand questions, as well as the insights held in all of human knowledge, could speed financial professionals’ learning process and give them a systematic way of approaching and answering questions. The result is quicker and more efficient decision-making.
But, as smart as computers are, they can’t teach themselves natural language—yet. The first step is training. So how do you train your computer? Think about it in the same way you teach a toddler to talk.
Early questions are as simple as possible with a concrete answer —“What is X? What happens to X if Y?” By repeatedly linking the pattern of words in a question with its answer, the computer learns enough about linguistics and the subject matter to understand a new question and make a good guess at a relevant answer. Some questions like “what role might politics play in stock market performance?” are answered well by recalling excerpts from a recent article, while other questions like “how much has the value of the pound declined this past year?” may require action or calculation. Through trial and error, it builds confidence in its ability to provide an answer based on a certain set of content. Gradually we increase the questions’ complexity. From currencies and bonds to global versus regional economic issues, by giving the computer diverse examples of good questions and answers in a consistent format, it begins to speak to you in the same way you would speak to your economist or financial strategist.
To be fair, training is an ongoing process with plenty of hurdles, as we can never fully explore all of the different questions that could be asked of the system. Just as humans are constantly learning new words and building connections between thoughts, words, actions and information; computers have to be brought up to speed with us. And since financial markets are constantly evolving, what might have been a good answer today might be the wrong answer tomorrow.
Still, our computers could be our avenue to the next big thing in a variety of industries, including financial services. In fact, we found that 49 percent of investment firms expect to be using innovations in AI, such as natural language processing, in the next five years. Meanwhile, 82 percent of customers say their investment provider will need to stay at the forefront of technology in the next five years, so there is still a lot of ground to cover. It’s worth helping computers learn a new, more robust language so we can all move forward together.
As head of Quantextual Research at State Street Global Exchange, Stephen Lawrence blends machine learning and big data with the contextual knowledge of human insight to streamline the investment research process. When not listening to U2, Steve listens to audio books on innovation, economics and the science fiction writings of William Gibson.