Data of the ocean, the human helm

Trading stocks, targeting advertising audiences, directing political propaganda, arranging appointments, defeating opponents in the TV quiz show "Dangerous Edge" (Jeopardy), and even choosing bra sizes: Computer algorithms are doing all this work and even covering more areas.

But behind the scenes, an old helper is undoubtedly playing a more and more important role, that is human.

Although the algorithms are becoming ever more powerful, fast and accurate, the computer itself is often only literal, and they often fail to understand the nuances of different contexts and meanings. Although these machines are so powerful, they are not always able to decipher the vagueness of human language and the mystery of human logic. Now, however, they are being asked to produce results that are more in line with human habits.

"While computers are smart, they can still be stupid," said Tom Carnegie, a computer scientist at Carnegie Mellon University (Mellon University). M. Mitchell (Tom M. Mitchell) said.

Thus, while programming experts are still writing step-by-step instructions for computer code, additional manpower is needed to do more detailed work, because the work being done by computers is becoming more complex. People have to evaluate, edit, or correct the work done by the algorithm. Alternatively, people would assemble a database of online knowledge and check and check it, essentially creating a memo that allows computers to quickly find answers. Humans can interpret and tune information into content that computers and others can understand.

The question-and-answer technology like Apple (http://www.aliyun.com/zixun/aggregation/5541.html ">apple" Siri) and IBM Watson, in particular, relies on this burgeoning human-computer collaboration. The algorithm alone is not enough.

Computer algorithms and engineers dominate Google's business and culture. But even in such companies, people are more involved in search results. Google needs human help in two ways. From a few months ago, if a user typed the name of a famous person or place, such as "Obama" or "New York", Google would display a summary of the information on the right side of the search results. These profiles are taken from databases that store knowledge, such as Wikipedia (Wikipedia), the CIA world profile, and Freebase, which acquired Freebase's parent company Metaweb in 2010. These databases are edited by people.

When Google's algorithm finds that a search keyword has a matching profile, the search engine crawls the information by command, not just the page link.

"Our ideas have changed," said Scott Huffman, the Google's technology director for search quality, Scott Heffman. Some of our information resources have been more organized by human beings. ”

Other human assistants have evaluators and reviewers who help Google fine-tune its search algorithms. Google's search algorithm is automatic and powerful, and can handle 100 billion search requests a month. "Our engineers are tweaking the algorithm, and real-person helpers will help us determine whether the proposed changes are really improving the program," said Huffman of Google. ”

Catherine Young, 23, is a Katherine of Google. She is a contract worker and a college student at Georgia Macon. She will see a vague search keyword, such as "What the King holds," and the results of two groups of Google searches. What she wants to do is to rate the relevance, accuracy, and quality of the results. In the search results of this imprecise keyword, some of the page links in front of it said that the king usually held a ceremonial scepter, which was a reasonable conclusion.

Yang said her judgment "is not completely black and white, some of which are subjective". She added, "You have to try to stand on the point of the person looking at this keyword." ”

IBM's Watson is a powerful question-and-answer computer that beat the winner of the dangerous edge programme two years ago, and people are training it recently to help doctors diagnose disease. But it also needs human help.

To help Watson prepare for the work of the Doctor, people were given the medical literature, scientific reports and digital cases that removed the patient's identity information. Instead of answering questions, Watson asked the clinical doctors and medical students at the Cleveland Medical Center (Cleveland clinic). They will give the answer through a function called "Teaching Watson" (Teach Watson) and correct the computer's errors.

FindTheBest is a fast-growing start-up in Santa Barbara, Calif., 25-Year-old Ben Thelles, a product manager for the company. The company claims to be a "contrast engine" that can find and contrast more than 100 subjects and products, ranging from universities to nursing homes, from smartphones to dog breeds. The website was launched in 2010, and the company now has 60 full-time employees.

Taylor helped design and edit the website's education section. He majored in English, not an engineer, but self-taught, and became an expert in exploiting hidden data from research in the education sector and other fields. His research methods include communicating with educators through conversation and email. He is an information detective.

On the FindTheBest Web site, information about more than 8500 colleges can be found quickly, depending on the geographical location, subject and tuition fees. Opening a university page, summaries, charts, and pictures will also show a wealth of information, even including the gender ratio and ethnic composition of all teachers and students.

Taylor and his team wrote the summaries and designed the original charts and pictures. For example, they would choose the most relevant to college students and their parents from hundreds of data on university fees. But most of their information is in the template, and the code that the computer can read is annotated. So the whole process has become more automated, and Taylor and others basically just give the algorithm a "crawl" command, and the algorithm executes.

Computer algorithms are improving, but the algorithm alone is not enough.

"It needs to be judged that the smaller data sets are the most important to be able to tell by feeling," Taylor said. "In order to do this, there must be some human involvement." ”

Translation: Cao Li, Linmonk

(Responsible editor: The good of the Legacy)