Will AI Ever Replace Translators? - Part 3, Ambiguity
To continue the discussion on why AI still has a lot to learn to be able to translate like a professional, I'd like to talk about ambiguity.
In linguistics, ambiguity can come from anything: phonetics, grammar, syntax, intonation, semantics, and even punctuation. It is usually classified into four types:
When working on a text, human translators can easily understand the correct meaning based on the context. Can AI do the same? Let's see.
We have this sentence translated into Romanian (I've chosen Romanian because I'm a native speaker and can confirm beyond doubt whether the translation is correct or not):
ChatGPT was not able to understand from the context that "English" refers only to poems. In its Romanian equivalent, Angela seems to like both English poems and English novels.
Additionally, ChatGPT made a major mistake in translating "Angela likes." In Romanian, when you're saying that someone likes something, you have to use the Dative case, while ChatGPT used… well, such a wrong form that there's no need to get into too many details.
DeepL, on the other hand, sensed this subtlety and used the Dative case correctly. Nonetheless, it could not decipher the ambiguity properly based on the context.
Here's another example:
Both ChatGPT and DeepL did not understand the first sentence correctly. The meaning of "Jake may not see it" is that Jake is not allowed to see it - after all, it's a cake for him, so it's a surprise, right?! AI translated it using exactly the same verb as in the second sentence (where it's correct). The meaning of the first sentence is lost.
To continue the discussion on why AI still has a lot to learn to be able to translate like a professional, I'd like to talk about ambiguity.
In linguistics, ambiguity can come from anything: phonetics, grammar, syntax, intonation, semantics, and even punctuation. It is usually classified into four types:
- Phonetic ambiguity, which is caused by words that have a similar pronunciation (flower (/flaʊər/) and flour (/flaʊər/)).
- Lexical ambiguity, which is caused by polysemous words.
- Syntactic ambiguity, which appears when the structure of a sentence can be analyzed in different ways.
- Pragmatic ambiguity: as Weiwei Dai puts it, "[p]ragmatics deals with specific utterances in specific contexts, with particular emphasis on the ways in which different social contexts influence language interpretation."
When working on a text, human translators can easily understand the correct meaning based on the context. Can AI do the same? Let's see.
We have this sentence translated into Romanian (I've chosen Romanian because I'm a native speaker and can confirm beyond doubt whether the translation is correct or not):
| English | ChatGPT Translation | DeepL Translation |
| Angela likes English poems and novels. She reads only French novels. | Angela îi plac poemele și romanele englezești. Ea citește doar romane franceze. | Angelei îi plac poeziile și romanele în limba engleză. Ea citește doar romane în limba franceză. |
ChatGPT was not able to understand from the context that "English" refers only to poems. In its Romanian equivalent, Angela seems to like both English poems and English novels.
Additionally, ChatGPT made a major mistake in translating "Angela likes." In Romanian, when you're saying that someone likes something, you have to use the Dative case, while ChatGPT used… well, such a wrong form that there's no need to get into too many details.
DeepL, on the other hand, sensed this subtlety and used the Dative case correctly. Nonetheless, it could not decipher the ambiguity properly based on the context.
Here's another example:
| English | ChatGPT Translation | DeepL Translation |
| I've made a cake for Jake. I'll show it to Hanna and Oliver but Jake may not see it. I've made a cake for Jake. I'll invite all his friends over, but Hanna may not come. | Am făcut un tort pentru Jake. O să i-l arăt lui Hanna și lui Oliver, dar Jake s-ar putea să nu-l vadă. Am făcut un tort pentru Jake. O să îi invit pe toți prietenii lui, dar Hanna s-ar putea să nu vină. | Am făcut un tort pentru Jake. O să-l arăt Hannei și lui Oliver, dar Jake s-ar putea să nu-l vadă. Am făcut un tort pentru Jake. Îi voi invita pe toți prietenii lui, dar Hanna s-ar putea să nu vină. |
Both ChatGPT and DeepL did not understand the first sentence correctly. The meaning of "Jake may not see it" is that Jake is not allowed to see it - after all, it's a cake for him, so it's a surprise, right?! AI translated it using exactly the same verb as in the second sentence (where it's correct). The meaning of the first sentence is lost.
It is able to assess the nuance of language, it would be good to see some examples of where this works and does not work.

Reactions: Olajumoke Bello