[featurebug] Beware of AI generated and suggested definition!

[gbonnema]

The newly introduced AI generated definition is shown at the top of alternatives in a brighter color than the other definition. It suggests being better than the others. If you have a new word current, then an “enter” will copy that suggestion.

Unfortunately it is often not better. Let me explain in an example. See screen prints below.

The Finnish word “kosketuksesta” is an inflection of the word “kosketus” and means “touch, (military) contact” according to wiktionary. As you can see in the screen prints it suggests the definition “contact with reality” which is a translation of the words “todellisuuden koskeutuksesta”. That is not a definition of the word! It is a chunk out of the translation.

How can beginners trust the preferred suggestion in more difficult languages?
Those are way harder to verify. Also, the “preferred suggestion” spreads! Every time someone presses enter on a new word.

Studying a new language is hard enough. It would be of more value if the AI generated suggestion would be verifiably correct.

I understand the inclination to use the newest of the newest gadgets, but without proper testing (by a knowledgeable language learning expert), they can easily make things worse.

Can you please correct this feature, fast? As the wrong definitions are spreading as we speak.

[davideroccato]

@gbonnema

That is not a definition of the word! It is a chunk out of the translation.

I second this. It wasn’t like this before. I don’t understand why AI definitions are now the translation of a short sentence and not focused on the single word!

[MarkSiip]

How would a software verify a translation is correct? Like an native speaker would translate it?

I’m learning Finnish which I could as an child, but forgot much of it, and now try to relearn it. But I still know and remember some things like a native speaker. And think in general the AI-translations are often quite good. Mostly much better then the none Ai-transalations which often are totally wrong.

[Obsttorte]

@MarkSiip

How would a software verify a translation is correct?

The software probably can’t, unless it is able to countercheck it with existing dictionaries. However, software developers can test their software, can’t they?!

[gbonnema]

If you are responding to my post, then you might be misunderstanding. I am not speaking about AI generated translations. That is a different subject. I am speaking about definitions. See the screenprint with the dictionaries on it (Kieli, Wiktionary, Kielitoimiston sanakirja … ).

Maybe reread the post and tell me if I am being unclear. It happens. The screen prints might help.

[MarkSiip]

Ok, properly I misunderstood you. But even when I read your first post, I don’t really understand your concern. But never mind.

[Obsttorte]

I may be able to shed some light here. The issue is, that when clicking on a single word one would expect to get a translation suggested for that very word. In the case @gbonnema tried to illustrate, when clicking on kosketuksesta he would expect to get either touch or (enemy) contact or anything similar in meaning that fits the context proposed as translation. Instead, the translation suggested is contact with reality, which is not the translation of kosetuksesta but instead of todellisuuden kosketuksesta. So the ai suggests a translation that incorporates other words of the sentence, which it shouldn’t.

My own experience thus far is that, besides this very behaviour, it suggests translations that don’t fit the word familiy (I click on a noun and the suggested translation is a verb, for example) and once I even had a translation that did not only incorporate the next word of the sentence, but also the translation was a mixture of German and English. Something I already encounter regularly when translating a complete lesson via ai (for some reason I seem to be the only user encountering this issue, or at least the only one who has reported this).

[MarkSiip]

Maybe the AI translation is better or worse depending the language?

I study the course " Finnish - Learn Finnish by Listening and Reading" and sometimes the sentence translation, in my case to Swedish, is quite different from a word to word translation. You must look at the whole sentence, in order to understand the meaning, and to be able to translate it. Translation of a specifik word can wrong in that context.

Despite how good or not the AI is, properly it will (never?) be as good at translating, as a native speaker, who knows both languages well.

[zoran]

Thanks for all your feedback. We are looking into improving the feature.

[ejackson]

As soon as I have some facility with a language, I start using a dictionary written in the language to understand the meaning of words just as I do in my native English. Once you have a basic vocabulary, singe word translations are just too unreliable, even sometimes useless, because the single word translation might not fit the particular context. That’s what’s happening in several of the posts above.

I did this with Norwegian when I first when I advanced to Intermediate according to LingQ. I’ve just done so with Spanish at about the same level of vocabulary. It’s slower, so I only do it right now for words where the LingQ translation doesn’t seem to fit.

I don’t really fault the translations or word suggestions in LingQ. I just use them as a guide or first approximation, then it’s up to me to dig into a problem word on my own with other plentiful resources. I actually learn more that way.

[jt512]

So far, the German AI definitions seem quite good to me. They seem to be translations for the word in the context of the sentence than standard dictionary definitions. In other words, if a word has several meanings, the AI translation sees to be the best one given the context of the sentence.

[LuckieNoob]

If you have multiple translations, just choose the best one. The context depended AI translations are really good imo. . In the past you had to live with the google translate translation or look up the word manually if there was no community suggested translation. With this new system at least it pulls a new translation everytime you encounter the word in a new sentence. So you can pull many different translations for the same word. Over time this will result in enough correct translations per word.

[LinVega]

they will definite rework their prompt on OPEN AI. There were many words in German that doesn’t exist or misspelling, then out of the thin air there is a translation from the context.

[LinVega]

by asking CHATGPT if this word is a legit word in a certain language first, return a bollean, then decide whether the translation can be generated or not

[Obsttorte]

@LuckieNoob

The context depended AI translations are really good imo. . In the past you had to live with the google translate translation or look up the word manually if there was no community suggested translation. With this new system at least it pulls a new translation everytime you encounter the word in a new sentence.

You misunderstood the issue. The main problem isn’t that the translations are wrong as part of the sentence, but that - as illustrated in the opening post - the translation often doesn’t translate the hilighted word solely, but rather a whole word group that word belongs to. So the translations appears to make sense, as it somewhat does so, but is still wrong in the sense that it includes much more then the hilighted word actually means.

It also makes it harder to distinguish between those cases where a single word indeed has a meaning for which to express you would need several words in you mother tongue and those cases where the meaning changes due to the usage in combination with other words (auxiliary verbs, for example) or due to context (if the word is used metaphorically, for example).

This gets worse the less skilled one is in the language he’s learning and the more distinct that language is from those he already knows.

[gmeyer]

I love the idea about using AI to assess whether or not a word is a real-world in a language.

It would be greatly if LingQ could handle 1) technical breadcrumbs such as snippets of URLs, parsing errors of HTML fragments as well as 2) proper nouns in context such as person names and place names.

For proper nouns, there is an element to learning them that LingQ could help in.

For instance, perhaps a good definition of “Müller” is a “common surname across German-speaking countries,” and then for “Huber,” such as a “common surname in Bavaria.” For “Alex” maybe a “common surname and nickname for both men and women.” To “know” names, doesn’t it mean to know such as: a) is it a surname, given name, and/or nick name ?, b) what area is it used, whether across the country, from a region or foreign?, c) is it common or uncommon, and maybe even d) what age group is it typically associated with.

I wonder if similar could be done for place names. Should a definition of, for instance, “Nürnberg” be “Nuremberg, second largest city in Bavaria” or even indicate “known for cultural heritage and economic importance.”

Learning a language is more than learning vocabulary and patterns of expression for more than what one already knows. It’s also about expanding what one knows. I think proper names are a key example.

While this is been outside the scope of concern of historic dictionaries, perhaps that’s a legacy blocker that doesn’t really make sense ahead in the age of AI.

To me, when “comprehensible input” and “AI” are combined, I wonder if this “just makes sense.”

[Obsttorte]

@gmeyer

For instance, perhaps a good definition of “Müller” is a “common surname across German-speaking countries,” and then for “Huber,” such as a “common surname in Bavaria.

Shouldn’t it be clear from the context whether the word in question is a name? In the examples given the definition could also be: “Müller” is a “person who works in a grain mill” or Huber “medieval name for a pawn who owned a Hube, a large area of land, granting him the status of a yeoman”.
If the definition is just “Müller” = “Miller”, “Schneider” = “Taylor” etc. this should suffice, don’t you think?! And how common a name is will be clear once you’ve read a lot. Common names will appear often.

I can understand the idea behind your proposal, but in the context of LingQ I rather have

1. already existing features work properly
2. flaws that exist, especially in the ui design, fixed
3. new features discussed, especially based on the feedback of the users

in that order.

It is probably a bit much to ask for LingQ to become a tourist guide introducing you to the culture of those nations the language you learn is spoken in. And I am not sure whether many users of LingQ tend to solely us this platform for their language aqcuisition (which for me, like for you, includes familiarizing with the culture). They surely rely on other sources, too.

[gmeyer]

@Obsttorte , I was thinking about that.

Indeed it’s clear that it’s a name, but I’ve found it helpful working in a globalized culture to understand both what given names are feminine vs masculine, what are common nicknames, and the region from which a surname comes.

That Muller is Miller is rather irrelevant in my daily life.

What’s more relevant, such as in understanding company politics, is to understand where names come from. For instance, when several persons come from the same region, there is a chance there a higher level of communications among them. And then there’s such as this, it aids for small talk. If I know someone comes from region A and that their sports team won a big match over the weekend, it matters. Similar with recent or upcoming regional holidays and their days off. At least a whole lot more than whether someone’s paternal ancestral namesake was a yeoman.

Names matter. Insight for small talk matters. Even with near indifference to broader cultural concepts, if you’re going to interact with people in the language being acquired, it’s, for instance, good to know that a name is a masculine given name vs feminine given name, before discussing them in a conversation.

Do I see any point in translating Schneider to Taylor? No.

Would I like to see LingQ handle proper names better? Yes. Even if just to map to “A surname” or nothing. It would be better if LingQ handled proper names like they do numbers–to pretty much ignore them.

Could LingQ think about helping more. Yes.

I was triggered by the prompt of what exactly are “real worlds” in a language. Technical snippets of such as URL, numbers, and proper nouns aren’t really.

But you’re absolutely right IMO to critique me on the crawl, walk, run of it.

Fixing basic bugs in core features has to be priority.

Every time I see a new video of Steve’s pop up on “why not to study grammar” or whatever about language acquisition, I anxiously wish he’d stop blogging on his hobby of language and start posting on best practices in PDLC and SDLC in a software company, from what I understand to be closer to his day job. I’m hoping improved software quality isn’ t too big of ask.

[Obsttorte]

@gmeyer Just to avoid any misunderstandings. I don’t neglect the importance of the aspects mentioned by you. Although I am not sure whether there is a necessity to show interest in something that wouldn’t interest you in your native environment, either (just as you are mentioning sports ) I am just not sure whether this is something in the scope of a software like LingQ, at least currently.

P.S.: My idea with proposing “taylor” as translation for “Schneider” was that it was a common practice in medieval times to use job descriptions as surnames. So an English speaker could guess that it is like that in German for example and therefore anything that is a job title can also potentially be a name (unless it is something like software developer, noone is called like that ). On the other side I am not sure whether one can really easely memorize names, as their original meaning isn’t really important in the context of a story, for example. So if it doesn’t occour very often, you may forget it beeing a name anyways.

[gmeyer]

It’s the mispronouncing and more so the forgetting of someone’s name that seems to be a rather universal faux pas.