A new study may explain why Google Translate, which was earlier revealed to have made errors in translating some words about China, has been published. The study suggests that government censorship, including that of the Chinese Communist Party, can influence artificial intelligence operations. According to the study, artificial intelligence algorithms can learn words and phrases that are under scrutiny and thus influence the language used in translation programs.
Margaret Roberts, a professor of political science at the University of California, San Diego, and doctoral student Eddie Yang studied AI language operations trained by two sources: the subjects of their study included the censored “Wikipedia” website in China. “The subjects included the Chinese version of Wikipedia, which is blocked in China, and the similar Chinese website Baidu, which is censored by the Chinese Communist Party.
Roberts and Eddie used “Wikipedia” and “Baidu” respectively to assess whether news items were “positive” or “negative” in meaning. The results show that when Wikipedia encounters the words “election,” “freedom,” and “democracy,” most of the entries are “positive” or “negative. The results show that Wikipedia mostly evaluates “positive” when it comes to the words “election”, “freedom” and “democracy”, while “Baidu” evaluates “positive” when it comes to the words “surveillance”, “social surveillance” and “democracy”. “social surveillance” and “Chinese Communist Party”. They also found that Wikipedia considers “democracy” to be associated with “stability,” but Baidu’s encyclopedia considers “democracy” to be associated with “chaos. “.
Roberts believes that the differences found in the study may not be entirely due to government censorship, but may also be the result of self-censorship by the writers themselves or the fact that they come from different cultures.
Roberts said it is important to recognize that government policies may lead to other forms of “bias” lurking in artificial intelligence systems. Roberts said, “We think this is a starting point for us to begin to try to understand how government-shaped data plays a role in the learning process of AI.”
The study will be presented at the Fairness Accountability and Transparency Conference (FAccT) 2021 in March.
Last month, a user found that typing “China breaks promise” into Google translate gave the result “China keeps its word”; however, typing “US breaks promise” will give the result of “the United States breaks promise”. However, the error will be fixed later.
Recent Comments