[ad_1]
The insight that drives CaliberAI is that this universe is a Bounded unlimited. Although AI moderation cannot conclusively determine truth and falsehood, it should be able to identify a subset of statements that may even constitute defamatory statements.
Carl Vogel, a professor of computational linguistics at Trinity College Dublin, helped CaliberAI build the model. For potentially defamatory statements, he has an effective formula: the statement must name an individual or group implicitly or explicitly; make a claim as a fact; and use certain taboo language or ideas, such as about theft, drunkenness or other Suggestions for misconduct. If you provide a sufficiently large text sample to a machine learning algorithm, it will detect patterns and associations between negative words based on the companies they retain. In this way, it can wisely guess which terms (if used for a specific group or individual) put a piece of content into a defamation danger zone.
Logically, there is no data set of defamatory material available for CaliberAI to use, because the publisher works very hard to avoid publishing these things to the world. Therefore, the company established its own company. Conor Brady first used his long experience in the press to generate a list of defamatory statements. He said: “We considered all the annoying things that can be said to anyone. We chopped them up, cut them and mixed them together until we could endure the fragile situation of the entire human race.” Then, by the team A group of annotators supervised by computational linguists and data linguists in Alan Reid and Abby Reynolds used the original list to build a larger annotator. They used this fictitious data set to train the AI to assign probability scores to sentences, ranging from 0 (definitely not slander) to 100 (call your lawyer).
So far, the results are similar to a defamatory spell check.You can play Demo version The company’s website warned: “When we improve the predictive model, you may notice false positives/negative situations.” I typed in “I believe John is a liar” and the program spit out 40 probabilities below the defamation threshold . Then I tried “everyone knows that John is a liar”, the program spit out 80% probability, marked “everyone knows” (fact statement), “John” (specific person) and “liar” (negative Language). Of course, this does not solve the problem. In real life, my legal risk depends on whether I can prove that John is indeed a liar.
The company’s chief technology officer, Paul Watson, said: “We are categorizing by language and returning the suggestion to our customers.” “Then our customers have to use their years of experience and say,’I Do you agree with this suggestion?’ I think this is a very important fact that we are building and trying to do. We are not trying to build a real engine for the universe.”
It’s fair to wonder if professional journalists really need an algorithm to warn them that they might slander someone. “Any good editor or producer, any experienced journalist, as long as he sees it, he should know it.” said Sam Terilli, a professor at the School of Communication at the University of Miami and the former general counsel of the University of Miami. Miami herald. “They should at least be able to identify those statements or paragraphs that may be risky and worthy of in-depth study.”
However, this ideal may not always be achievable, especially in a period of limited budget and high pressure, requiring publication as soon as possible.
“I think news organizations have a very interesting use case,” said Amy Kristin Sanders, a media lawyer and professor of journalism at the University of Texas. She pointed out that reporting breaking news is especially risky when the news may not go through a thorough editing process. “For small and medium newsrooms-there is no general counsel to accompany them every day, they may rely on a large number of freelancers, and there may be a shortage of staff, so there are fewer and fewer comments on editorials. Compared with the past, I do think these tools may valuable.”
[ad_2]
Source link
