104. Read the following passage, then answer the question below.

The idea of translating languages by machine is older than the modern computer. In 1949, the American mathematician Warren Weaver suggested that a foreign text could be treated as a code to be deciphered, and his memorandum encouraged researchers to try. In 1954, a demonstration by Georgetown University and IBM translated about sixty Russian sentences into English and generated enormous excitement. Journalists predicted that the problem would be solved within a few years.

It was not. The early systems relied on dictionaries and hand-written grammatical rules, and they stumbled over ambiguity. A word like 'bank' has several meanings, and a rule that picks the right one in one sentence fails in another. By 1966, a committee in the United States concluded that machine translation was slower, less accurate, and more expensive than human translation, and funding dried up for years.

The field revived in the 1990s with statistical methods. Instead of writing rules, researchers fed computers millions of sentences that humans had already translated and let them calculate which phrases most probably corresponded. The results were better, though often awkward. A more dramatic leap came in the 2010s with neural networks, which learn to represent the meaning of a whole sentence rather than translating phrase by phrase. Translations became noticeably more fluent, and services now handle dozens of languages instantly.

Still, fluency can conceal errors. A neural system sometimes produces a smooth sentence that quietly changes the meaning or invents a detail, and it performs poorly on languages with few available texts. For this reason, professional translators have not disappeared. Many now work alongside machines, correcting drafts rather than starting from nothing, and the human skill of judging what a text is meant to do remains hard to automate.

The word 'them' in paragraph 3 refers to