Hi HN, I'm Musab and I'm building Manalect. Find TL;DR at the end.

Most language apps fall short for me in two ways:

1. They’re not personalized enough. I get taught things I don’t care about, things that are too hard for me, and things I already know. 2. They don’t teach that well. I keep seeing question types where my brain does almost no work to figure out the answer, therefore learn nothing.

I think the biggest reason they're not personalized enough is that they don't have a detailed enough representation of a language. Most apps have efficient review scheduling for the words you learn. But if you make the mistake of using "play" instead of "played" they're not able to identify that as a component that should have its own schedule for review. And if you don't track that, it essentially depends on luck whether a person reviews that information before forgetting it again.

Essentially language isn't just made up of words, but also different forms, conjugations, phrases, grammar, and a lot more components.

But mapping out the basic units in full detail isn't enough. It's also very important to know how those basic units relate to each other. Because of course some units share a lot in common, and their relationships form another unit that needs to be tracked.

Take Spanish: “el gato negro” (the black cat) but “la casa negra” (the black house). “negro” becomes “negra” because “casa” is a feminine word. That rule isn’t part of any one word. It links the noun, its article and the adjective, and it shows up in almost every sentence. So it needs its own review schedule too. And it relies on a person knowing those things that make it up to be able to build that block.

That was briefly about how I'm working on solving the first failure mode. Second is simpler but also very interesting. We know that for you to learn something, you need to experience difficulty. This is called desirable difficulty.

For this reason, some question types, like multiple choice, are known to be inefficient. But that's not fully true. The reason multiple-choice questions can be inefficient is that usually the alternative options are so much more different from the correct one that it is very easy to guess which one is the right one. However, if you make a simple change where you make it so the alternatives are a lot closer to the correct answer, the person has to put more effort into recalling the correct option. In research, this is known as competitive distractors.

And of course, desirable difficulty is not the same for everyone. This is quite simple to imagine. Even games allow you to change the difficulty, and the difficulty that is most efficient for learning doesn't necessarily have to be the difficulty a person can sustain for the long term. So it's also very personal.

I had written a lot more about how it works in detail, but due to the allowed length for this post, I'll get to how specifically the app works on an application level. Please let me know if you have any questions. I'm very passionate about this, and I would love to talk more.

How it works:

1. Translate sentences, define words, chat with AI, or import an article or a song 2. The content gets broken into granular components, not just words but grammar and more complex building blocks) and saved for you 3. You study them as short tests in real sentences 4. Stuff you already know gets skipped, and new stuff builds on what you know

You can try it at manalect.com, no sign-up needed. Swedish, German, French, Spanish, Portuguese, Polish or English. It's free with a daily limit.

It's at an early stage, and there might be bugs, but I'd love to hear from you about them so that I can fix them right away.

I'd also love to hear your thoughts, and how you learn languages in general. I'll be around all day to answer questions.

Show HN: Google translate replacement for language learners