Seeking feedback on BETA FEATURE: Chat GPT individual word translations on iOS

We’ve implemented an experimental feature that can be enabled from the iOS Reader Settings, in which you can get individual word translations from Chat GPT.

There are two different settings

  1. Contextual Hints: These translate the individual word in the context of the sentence.
  2. GPT Hints: These are just straight up word translations that don’t take the sentence into context.

When you have either of these settings enabled, these will be the top hints shown when you click on a blue word. The hint will have a small “AI” shown alongside it indicate it’s an AI translation.

We would love to get feedback from iOS users.

5 Likes

This sounds like a really useful feature. Are there plans for it to be available on the web version (or Android) as well?

1 Like

This sounds great and I definitely appreciate our involvement, however, I would have further questions:

Will we be able to see both chatgpt hints (both of them), the popular meaning, and the dictionaries as it is right now?

How we will be able to select those AI hints and make them our own if we want to?

In which languages will be those AI hints? Native languages we are reading, or our own native language, or both?

Thanks.

1 Like

@Ryanaissance yes it will eventually be available on all platforms. We are building an API to feed to all platforms. But in the meantime, it’s been built on the front end for iOS.

1 Like

@davideroccato

You will see both types of chat gpt hints if you enable both settings. You will also see popular meanings below. And the dictionaries are still there - no change in that regard.

To select the AI hint, first enable the setting, then click on a blue word (or click a yellow word and swipe up on the popup). The top meaning will be the AI meaning - when you expand the popup, it will show an “AI” label. Select that hint.

The language of the hints will be in the selected dictionary language. I’m not sure exactly what languages are available right now. But I will confirm on that shortly.

1 Like

Oh, I get it, it’s already in the settings.

@tommy_learner_success I only see Contextual Hints, which I have enabled. I don’t see GPT Hints, where is suppose to be?

First problem I have. Now, it gives even less room to the popular meaning. Because one is taken from the AI. Isn’t it possible to add a sort of arrow at the bottom to EXPAND all popular meaning, without having to select one?
We see the hints, if we want to, we select one, otherwise we expand all hints with an arrow, and then we select one, or two, or more, before closing the window.

EDIT: I see I can open the popup by swiping up instead of having an arrow. I forgot about it, there is no clue about it.

1 Like

I’ve been using this feature for a month or so during the alpha and it’s pretty cool. Created more than 10k definitions thus far. Especially helpful in smaller languages that see little traffic and have very few available definitions. For example in Romanian, the difference between this and the auto-provided Google Translate or DeepL is literally night and day. In larger languages with many available dictionaries and definitions, like French or Spanish, this is feature is not really needed.
One problem of course is the reliability of a large language model. They can hallucinate and confidently give you completely confabulated nonsense definitions. But this is is really hard to counter by modifying the prompt alone (at least I was unsuccessful). I’ve been thinking, maybe LingQ could keep a real dictionary server-side and provide the model with these definitions (if available). This could potentially provide some “grounding” and make hallucinations less likely? That way the model would have something to go by, even if it actually has no clue. They never admit to having no clue in my experience.
Another thing that has been impressive is the way one can get it to conform to a specific style, e.g. using semicolons or slashed or newline separated etc. The rarely breaks out of the specified format.
Other problems are expected in this early stage, as the prompt is in English at the moment, it doesn’t really work well if I set my dictionary language to German. The prompt also instructs the model to provide the definitions in lower case, which is not ideal in German.

Eventually, I would suggest to try to get ChatGPT to output structured data like JSON, once this is closer to production. It’s cool to be able to play with the prompt on my device, but I’ve definitely had the model produce a novel-length apology letter once or twice, those now reside somewhere among the popular meanings. I’ve heard that Pydantic Models to validate the data from ChatGPT are a popular choice.

Thanks again for the great work, I had almost stopped using LingQ until this feature came around.

3 Likes

@tommy_learner_success It would be great, after expanding the popup, to have the entire definition in the native language as well, maybe after the Notes field.

It doesn’t seem to work by selecting 2 words and more.

1 Like

It’s a good idea. Compared with the popular meaning, that they hallucinate already, chatGPT would handle it better. Plus, with contextual reference, plus GPT hints that I don’t see yet, plus the popular meaning, there is more probability to get it right.

3 Likes

I was thinking a single monolingual dictionary per language might suffice, because ChatGPT (hopefully) knows the language and is able to translated it internally into the desired target language.
To be on the safe side legally, LingQ could start with Wiktionary, which is under a liberal license and can be conveniently downloaded. (I think we even discussed how to extract the data from Wiki dumps on this forum). There are also good bilingual alternatives for example in Chinese mdbg.net is very good.
Getting the publishers of “proper” professionally curated dictionaries to license their content to LingQ, is probably not going to happen, even (or especially?) if it goes through a LLM before users see it.

2 Likes

@bamboozled very good points you make, We are working on server-side translations to provide reliable data for the LLM.

1 Like

@davideroccato by entire definition, you mean all the different translations of the word, just not in that particular context?

And yes, unfortunately, it’s not yet set up for phrase lingqs. Just single words.

1 Like

As for which languages this is available, it’s available for all study languages and dictionary languages…but it works less well for less common languages. There’s just less data for the model to work with in less popular languages.

1 Like

@tommy_learner_success

  1. for entire definition I meant the meaning of the word. BUT not in the hints, but as extra expandable, like a Notes field. Same thing when you tap the word on Google and you have a sentence that explain the meaning. And of course, if there are more meanings, you have (1)… (2)…, and so on. If this is too complex, at least the meaning for the word in context, but it should be clear that it is contextual.

  2. where is the setting for GPT Hints? I only have the contextual so far.

  3. sometimes there is a big delay before the AI meaning appears.

  4. I selected a hint, and modified the meaning by adding something more in the Saved Meanings. I have noticed that the AI field below has been modified as well. I believe I have also accepted another hint that was already modified by another user, because there was much more information than usual. Is that possible? See screenshot below.

EDIT: I have done another one.

1 Like

To add to this, I always thought about Wikipedia as well. I use it a lot for a lot of terminology that are historical, or have other connotations.
ChatGPT can extrapolate from Wikipedia as well, if the word is not present in Wiktionary, and return a summarize meaning below 250 characters, which is the current limit!
It could be added to the Notes if requested.

1 Like

@davideroccato

  1. That’s a good idea. We’ll be exploring more ideas like this in the near future - how to use AI to provide more insights in the reader
  2. My apologies - GPT Hints is actually not available. Only contextual.
  3. Thanks, we’ll look into it. This is a beta feature, we will improve it.
  4. Yeah that sounds like what is happening.
1 Like

From the engineering team: An external database to train the LLM is unreliable for these context-based meanings. So instead, we will be building our own database/dictionary. We will be tracking how often the contextual dictionary entry is used. If it’s used a lot, we will recycle it so the LLM doesn’t have to generate a new one (which will make things faster).

Regarding the hallucinations of the AI output ( @bamboozled ), this is because this feature is currently using GPT-3.5. We will be improving the experience with more polished LLMs.

1 Like

We greatly appreciate the feedback here. Step 1 is simply to improve upon the popular meanings. In my experience, these contextual hints are typically better than the top popular meaning which shows. But if you guys find that not to be the case, please let me know.

2 Likes

@tommy_learner_success I will use it more from now on. From my point of view, I need to add some extra content, because it’s true that the meanings are more relatable, but they’re not really accurate compared to a real dictionary. They fit the sentence at the moment, but that doesn’t mean that the same translation will be good the next time we encounter the same word. That’s why I often create my own definitions.

But it’s good to have the AI option as well, to quickly accept it on the phone when no other options are available. Or to quickly check with other options to compare and accept a closer match.

Relying on how many times people accept a translation might not be a good idea, because we often accept translations just because we’re in a hurry or don’t want to bother too much. The same thing happened before, and there were a lot of popular meaning incorrect translations that were accepted by users anyway.

Maybe you can give a different “weight” to “Pro users”, and casual users, when they choose translations. Then you calculate on your own your priorities.

1 Like

It can translate to wrong language))
Even that are not in the list of dictionaries.
I learn German and have Russian and English translations, but ChatGPT suggested Spanish. I had to overload a lesson, and after it all was ok

1 Like