Korean has a perfectly phonetic written language. It was invented by King Sejong and his scholars in 1444 specifically to be phonetic. Koreans probably use “Tyranasaurus” and “tiramisu” pronounced as-is, and the translator app translated the portmanteau phonetically to English.
To add to that, credit where credit is due, LLMs can often pick up on things like this. Machine translation has been LLM-based (or some primitive ancestors of LLM) for many years even before the AI boom. So AI probably helped a bit here.
That’s my wild guess. I wouldn’t call it a hypothesis, I’m just talking out of my ass.
There are other usages in computer linguistics. My master thesis was a neural parser. Other usages are in pattern recognition in medicine for example. But your point stands that often it makes things worse
I had heard about the medicine thing actually. When the use case actually lines up with what it is, it makes sense as a tool. It’s that old adage though “When you have a hammer, everything looks like a nail.”
Is there any way I can read your thesis? I’m casually curious, and also have no idea if college thesis are allowed to be shared online with rando people like me.
Well, it parses natural language. In linguistics, or syntax to be precise, there are different ideas on how to build syntax trees. The most common is Dependency Grammar, basically just a tree where every word points to the word it refers to (the adjective to the noun, the subject and the object to the verb, the verb is the root). I applied this to a different syntax theory called Role and Reference Grammar. You can google the latter, if you want to look into neural parsers in general, stanfordNLP has modules for python and I think online tools as well and stuff.
Korean has a perfectly phonetic written language. It was invented by King Sejong and his scholars in 1444 specifically to be phonetic. Koreans probably use “Tyranasaurus” and “tiramisu” pronounced as-is, and the translator app translated the portmanteau phonetically to English.
That’s my hypothesis.
To add to that, credit where credit is due, LLMs can often pick up on things like this. Machine translation has been LLM-based (or some primitive ancestors of LLM) for many years even before the AI boom. So AI probably helped a bit here.
That’s my wild guess. I wouldn’t call it a hypothesis, I’m just talking out of my ass.
A hypothesis is basically a guess based on logical assumptions so you are there already.
Translation might be the only thing they genuinely do better than older tools.
There are other usages in computer linguistics. My master thesis was a neural parser. Other usages are in pattern recognition in medicine for example. But your point stands that often it makes things worse
I had heard about the medicine thing actually. When the use case actually lines up with what it is, it makes sense as a tool. It’s that old adage though “When you have a hammer, everything looks like a nail.”
Is there any way I can read your thesis? I’m casually curious, and also have no idea if college thesis are allowed to be shared online with rando people like me.
It depends in part in your ability to read German 😅 I wrote another comment elaborating a little and giving clues for “further reading”
Thats super cool! What sort of things did your neural parser do?
Well, it parses natural language. In linguistics, or syntax to be precise, there are different ideas on how to build syntax trees. The most common is Dependency Grammar, basically just a tree where every word points to the word it refers to (the adjective to the noun, the subject and the object to the verb, the verb is the root). I applied this to a different syntax theory called Role and Reference Grammar. You can google the latter, if you want to look into neural parsers in general, stanfordNLP has modules for python and I think online tools as well and stuff.