Three broad approaches
Translation software has changed in method more than once. Historians and engineers commonly divide it into three overlapping eras: rule-based systems, statistical systems and neural systems. The boundaries are blurry, because many products combined methods, and different groups moved at different speeds. The early optimism and the setbacks of the 1950s and 1960s are covered elsewhere, so this article starts after that period and focuses on how the methods differ.
Rule-based systems
From the 1960s through the 1980s and beyond, most translation programs relied on linguistic rules written by people. Typically a system analyzed a sentence in the source language, worked out its grammar, and then generated a sentence in the target language. Some used an intermediate step, in which the sentence was converted into a more abstract representation before being rendered in the other language.
These systems were built with large bilingual dictionaries and detailed grammar rules. A few were put into practical use, notably in settings with controlled vocabulary such as weather bulletins or technical manuals, where the range of sentences is narrow. Some commercial and government systems from this period were used for years, usually with human editors checking the output.
What they did well: Output in a narrow field could be fairly consistent and predictable, and errors could often be traced to a specific rule. What they did poorly: Building and maintaining the rules took enormous expert effort, and ordinary text with its idioms and exceptions often broke them.
Statistical systems
In the late 1980s and 1990s, researchers began to try a different idea. Instead of writing rules, they let programs learn patterns from large collections of text that had already been translated by humans, such as official records published in several languages. A statistical model estimated which target-language words and phrases were most likely to correspond to the source text.
This approach became dominant in the 2000s, helped by growth in computing power and in the amount of text available on the web. It powered the first free online translation tools that millions of people used. Rather than analyzing grammar in a traditional sense, systems typically matched and rearranged chunks of phrases.
What they did well: Coverage of many languages was possible without a team of linguists for each pair, and output was often more fluent than word-by-word rules. What they did poorly: Quality depended on how much parallel text existed, so languages with less data fared worse, and long-distance grammar or word order could still go wrong. Results could read smoothly and still be inaccurate.
Neural systems
Around the mid-2010s, translation systems based on neural networks began to replace the statistical phrase models in many major services. These systems learn a representation of whole sentences from large training sets and generate the translation word by word, taking wider context into account.
Companies and research groups reported noticeable improvements in fluency, and many users saw the change in everyday tools. The basic idea of neural networks is much older than this period, but the combination of more data, faster hardware and better training methods made it practical for translation.
What they did well: Output tends to be more natural, handling context across a sentence better than earlier phrase-based methods. What they did poorly: They can produce fluent text that is wrong, may invent or omit content, and still struggle with rare words, specialized vocabulary and languages with limited data. Judging quality also remains difficult, since automatic scores do not always match human judgments.
Other developments along the way
Several supporting tools grew up beside the main methods. Translation memory software stores segments that a human translator has already done and offers them again when similar text appears, which became widely used among professional translators. Speech recognition added spoken input to some tools, and software for handling character sets made it easier to work with many scripts.
Another notable shift was the arrival of translation as a free service in web browsers and phones. This made automatic translation an everyday way to get the gist of a page, even when it was not accurate enough for legal, medical or other sensitive uses.
What has and has not changed
The quality of machine output has clearly improved since the first experiments, and for common language pairs and everyday text it is often good enough to be useful. Still, the core difficulties that appeared in the earliest projects, such as ambiguity, idiom and the need for background knowledge, have not disappeared. They are handled better, not eliminated.
Human translators remain important wherever accuracy matters. The role has shifted for many of them toward checking and editing machine output, but how much that changes depends on the field. It is wise to treat any single claim that translation is “solved” with caution.