1. Forum
  2. >
  3. Topic: Spanish
  4. >
  5. There is something very wrong…


There is something very wrong with how the "Strengthen Skills" algorithms work.

So I start my daily Duolingo lessons for Spanisih. I see I have three decayed lessons - Adj. 2, Feelings, and Nature. No problem. It is a bit much for a decay for one day but I'm used to Spanish being overly harsh on this. But its Saturday, and I have time.

I figure I'll use the overarching Strengthen Skills option to try to knock out all three decayed lessons. After one round, Adj. 2 is back to gold. Excellent!

So I do another Strengthen Skills - but don't get a single question relating to Nature or Feelings. OK. I guess I have other words that need strengthening, I'll just run through it again.

About ten Strengthen Skills reviews later...and not a single Feelings or Nature quesion has come up. I know at this point the logical thing would be to just do both lessons individually, but at this point I feel it is a matter of principle!

Why won't Strengthen Skills pick up the words Duolingo feels I am lacking in Feelings and Nature? If my words are really that decayed in those lessons, why wouldn't Strengthen Skills pick those up? How many Strengthen Skills lessons will I have to do before it decides to strengthen what Duolingo has deemed to be my weakest areas?

This is seriously bizarre.

Edit: round 11 or so, Feelings has finally been regilded.

Edit 2: I have now earned over 250 XP points in Strengthen Skills, and have literally not had a single question come from my last ungilded lesson, Nature.

I will take a break from this experiment and try again later today. Will getting 500 XP points finally knock out whatever word or words in Nature I am lacking? We will see!

November 15, 2014



It could be broken, it could be math.

This is all purely one way you could do things, and I'm not saying it's how DuoLingo does it.

Each word has a value of how much you've used it, from 1 (hardly) to 10 (a whole lot). Let's say that a word needs to be strengthen when it's value drops to 3; and a skill needs to be strengthened when it has 3 words that need to be strengthened. And, since the trigger for losing gold is 3 words, we aim to strengthen at least 3 words per lesson. (I'm using small numbers to keep down the typing.) Now, the choice needs to be made: if you strengthen only one word out of the three, should the skill then become golden? If you pick yes, then you get back to gold quickly, but drop back down just as quickly--so, I'm picking that you have to clear all words to get back to gold. This is a key decision.

[Uppercase letters are Skills, lowercase letters are words.]

We start out with the following Skills: A-a1,b3 B-c2, D-d1,e1,f3 M-g3,h1 P-i2 R-j1

Now, only skill D has lost gold. If you picked Strengthen D, you'd get words d, e, and f. This would bring D back to gold, assuming you got through without losing any hearts. Which we will assume, as there's already too much typing!

However, if you pick Strengthen Skills, it will pick among the lowest level words; say you'd get a, d, and h. You've strengthened one word in D, but we need to clear all the words to get back to gold, so it's icon doesn't change.

Now we have: A-b3 B-c2 D-e1,f3 M-g3,h1 P-i2 R-j1

Doing a Strengthen Skills again, we pick the lowest values and get words e, h, and j. Still haven't cleared all of D.

Now we have: A-b3, B-c2, D-f3, M-g3 P-i2

Doing a Strengthen Skills again, we pick the lowest values and get words i, c, and b. Still haven't cleared all of D.

Now we have: D-f3, M-g3

You've done three passes though and almost all your words are good; but, due to the choice about how to handle a Skill changing back to gold, you're still showing D as needing strengthening, even though it's in exactly the same state as M--which is still gold.


Its strange that you seem to be saying that the Spanish tree is harder to keep golden as mine is my best behaved tree. It has taken far less work to keep it gold than either the French or German trees.

It seems that if we want golden trees the skills need to addressed individually. I too have had days where a string of strengthens regilds no skills, or only 1 or 2 out of 8 or so. My tactic now is to regild individual skills and then hopefully do a general strengthen or two, time permitting. Spanish has no skills decayed some mornings so then I go straight to the strengthening. Then I try to do three, that seems to be the minimum number to give it a chance to stay gold.


I wonder too. It usually just focuses on one thing.


Yup! My only ungilded lesson is Nature now, I do Strengthen Skills (I'm stubbon), and I get a lesson on clothes. Oh, Duolingo.


I have lots of skill to strengthen, Verb Present 3, Verb infinitive, countries, adjectives, pronouns, directions, education, abstract objects 1, feelings, present perfect, past perfect, nature, abs. object 2, v infinite 2, arts, and politics. lol


I didn't notice that, but sometimes two lessons are immediately "re"gilded when I do one "Strengthen Skills" and sometimes one single lesson does not even move forward one grade.


Even when you strengthen one skill another skill up the tree can gain points too and be regilded. (That's why it could be a clever tactic to start regilding down in the tree if your goal is to regild. If you really want to learn then the other way round is better.)


The problem here is that taking the full regilding for individual lessons isn't teaching me anything. It is just needlessly slowing me down. Which is why I would hope that Strengthen Skills would be a good way to strengthen words in many lesssons at once.

I am just about at the end of patience with the Spanish course, to be honest. I had hoped it would be a good review, but there are so many odd quirks with the tree that I don't get in other Duolingo languages. It is frustrating.


I don't see how individual strengthening is much different to a centralised algorithm that takes the state of the skills into account. If any, strengthening each single one and thus doing more lessons per skill should teach you more. A strengthen skill algorithm that would take this into account would be a little bit more convenient but, well, spotting the decayed lessons and indidually is not that hard. Why should this be slower than an algorithm that selects the lessons automatically?


Part of the issue here is that I am trying to manage three trees, and I am not made entirely of time. Also, I am fairly well versed in Spanish as is - this is more of freshening up/attempting to get into the later, more complicated grammar.

So the ability to keep regilding things as they come along in a more hollistic fashion is far more logical for me. When I tried to regiild each lesson individuially, I found that I spent all my time doing that, instead of moving forward. The only reason I'm spending all this time is because it is Saturday morning.

I'm not trying to sound whiney or complainy - this is a huge task Duolingo and the variosu teams are undertaking. I just find it a bit baffling that over 10-15 lessons of Strengthen Skills wouldn't strengthen the one Skill that the course wants me to Strengthen.

That to me is a major problem - If I wanted to, say, do a lesson or two of Strengthen Skills and one or two new lessons, then I am seeminly wasting time, because there is no apparent way to know if Strengthen Skills is actually doing what it purports to do. In the end I'll just end up with new ungilded lessons regardless.

I guess my concern is that if Strengthen Skills can't hit on my most decayed words, then it is ultimately not reliable enough to use, and I'll be back to trying to keep the entire tree gilded individually, which will mean that I will just have to quit, because the frustration that causes outweighs any benefit.

In short: Strengthen Skills in German and Irish seem to do a far better job at both keeping the tree mostly gilded and in re-teaching me words. I am just baffled as to why it works so differently/poorly with Spanish.


Yes, I have encounted the double gild before. Which makes sense - I would assume Strengthen Skills would attempt to pick up words from all the ungilded lessons. Whichever need the most strengthening. So sometimes it works well and sometimes it fails spectacularly.


Yes, but sometimes there is no progress at all in any of the lessons. I payed attention to this: sometimes you don't gain anything. :-/


I tend to switch between the strengthen skills exercise within a skill set and the strengthen skills button on the home page.
The home page strengthening exercises seem to be a bit harder and introduce new sentences using previously learned concepts. The skill page strengthening exercises are a bit more focused. They have variety from across all of the lessons of that skill. They introduce some new sentences, but fewer than the home page set. And of course, it brings up the strength directly in that skill.


This is how I do it in my other trees, and it works well. I'll do a lesson of Strengthen, a new lesson, and if theres a ungilded lesson I'll do that. But with Spanish I seemingly get 2 or 3 new ungilded lessons despite doing Strengthen lessons regularly. Which makes me worry that Strengthen Skills in Spanish is broken - which is further evidenced by racking up over 250 xp points today, and not a single whiff of the most decayed lesson I have.


This isn't an issue of something being wrong with the functionality so much as a badly named link which causes confusion and gives users an inaccurate expectation of its behavior. This is the result of a design decision made when the UI was revamped at the start of this year. The staff designer at the time (he's no longer with the company), said that the desire was to describe things at the "skill level" rather than at the "word level," even though much of Duo's core functionality still works at the "word level." In spite of its poorly chosen name, the general "Strengthen Skills" option does not target weakened skills, it targets the weakest words. The bottom line is that if you want to improve a specific skill, you must use the skill-specific strengthen exercise on the skill overview page.

There are some important things to understand here. First, the strength of a given skill is actually the average strength of the words taught within that skill. When a skill loses its golden status, it means that on average the words in that skill have dropped below 80% of full strength (skill strengths are represented having from 1 to 5 bars of strength, 1 bar means the average is between 0 and 20%, 5 bars it's over 80%, etc.). Second, because we're dealing with averages rather than specific values, there can be some hard to predict results. Is a skill degraded because all of its words are around 78% or because a relative few have degraded to 0%, drastically lowering the skill's average? This means that there can be a relatively large number of very weak words even when the average strengths of most skills are high enough to keep them from appearing to be degraded. Finally, the selection algorithm that chooses which words are presented in an exercise doesn't seem to chose the weakest words; instead, it appears to choose semi-randomly from among the weakest ~30 words, and there are some words that never show up at all for whatever reason (actriz in Spanish is one example that comes to mind).

What your experience demonstrates is that Adjectives 2 had enough words degraded sufficiently to get selected early on in your set of practice sessions. Nature and Feelings, however, do not have enough words with sufficiently strengths to be selected for a general "Strengthen Skills" exercise, and at least 90 words (9 words per session * 10 lessons) have a lower strength than the weakest words from either of those skills. This isn't at all unexpected given the behaviors that I described above. If you look at your words list sorted by strength, I suspect you'll find that those two skills' constituent words are not among the weakest two or three dozen word forms.

Here are a couple of older posts in which I explain how practice sessions affect word strengths and how decay rates rates work. Both of these things also factor into what you're experiencing a bit.



That's very interesting, thank you.

I have been having a similar issue with my Italian tree. When I went overseas it was all gold. I got back a week later, and I think around 10-12 skills had degilded. I agreed to do an experiment with Greg Hullender, and do two general strengthens per day, to see how long it took to regild. Now, almost two months later, more than 40 skills have degilded. Every day I expect it to turn the corner and start strengthening. It makes one wonder how it is set up, when that much consistent work is still resulting in a downhill slide. I would have expected that I would have enough of a spread of weak words across the tree skills to enable those to be picked up, resulting in an improvement by now. That still hasn't happened.

To contrast with that, my reverse tree, which I tested through entirely in one day, from start to finish, currently has 16 skills degilded. I had expected this one to fail catastrophically by now. I suppose, although I had no record of ongoing work on this tree, so no 'strong' words, at least all of the words have been seen within the last three months (I tested through it around 2 and a half months ago), whereas some words in my main tree have not come up for more than five months. I am also doing three strengthens per day on my reverse tree instead of two, as I had anticipated a lot more problems.

I understand how the OP is having problems, when one or two skills are weak, but the weak words chosen are not helping regild those. You would have thought though, that with the number of weak skills that I now have, the formula should work. I have been questioning recently whether there is a genuine problem with it. I have decided to give it another month, and if that doesn't improve it I am going to abandon the experiment, and possibly my tree, entirely.


This is a nice explanation of the way "Strengthen Skills" currently works. However, I think it ought to work more like the poster was expecting. I really don't enoying reviewing individual lessons that I've already completed. They can be repetitive and monotonous the first time around, so a repeat is really no fun.

I wish "Strengthen Skills" would work at the word level, with added weight being given to words belonging to lessons in need of review (the greater the need, the more the weight). I'd like it to feel more like a grab bag of sentences on various topics than a review of one or two topics that I may already be bored with.


I don't necessarily disagree. To be clear, though, my post shouldn't be taken as a defense of Duo's design decisions, or as an endorsement. How we think it should work is fair game for debate, but I was merely explaining how it does work currently, based upon lengthy observation, experimentation and explanation from devs over the last couple of years.


Right, I got that from your post. You gave a very clear explanation. I was just throwing in my opinion, for what it's worth. I know it's tough to try and please everybody without overloading the system with too many options.


Cool. I just wanted to make doubly sure my position on the matter was clear hehe.


Great explanation.

At first I also assumed Strengthen Skills would actually strengthen my skills so that I would not have to micromanage myself what needs reviewing. But it didn't take long to see Duolingo does not work like that. An extreme case are the extra lessons (e.g. Christmas, Flirting) which I have never seen in Strengthen Skills. There are also some words from regular lessons which I fail on purpose to see if they get repeated in the near future: some do but some disappear completely from review.

  • 2687

Most comprehensive reply ever! This should go somewhere on the (official) FAQ page!


You always have a wonderful way of clearly explaining things! Thank you! I think you hit the nail of the head! DL has a mixture of vocab and grammar lessons yet most of it's metrics are around vocabulary (words). I've noticed that the words learned in the vocabulary lessons are not well integrated in the grammar lessons resulting in the vocabulary exercises decaying faster than say, the past subjunctive exercises. Many of the more advanced grammar topics have 1 - 3 lessons and are quite repetitive using almost no words from, say nature or food or sports, etc.

I have run across a few sentences which incorporate topics beyond the vocabulary exercise I'm '''strengthening'' but not nearly enough. This is one area that I think DL could significantly upgrade course quality. I would like to see, say a mix of conditional perfect, subjunctive, imperfect, etc. sentences in the vocabulary exercises. The course should know (well the programmer should be able to tell) where each user is in the course is and incorporate all completed lessons into ''Strength Skills''. Vocabulary used in different contexts can, at times, be more difficult to remember or even take on a different meaning and there are more things to distract you (Is the subjunctive triggered here?)

Today it's Places and Nature. Business has turned up 3x in the last 5 days even though I have used the strengthen skills for business each time it has ''ungolded''. I'm guessing its because business words are integrated into very few of the other exercises outside of Business. The bottom third of my tree hasn't decayed in weeks!

It's a bit disheartening that you, and others, are explaining how Strength Skills works and not Duolingo staff. They should know exactly how it works and communicate it to users. In my opinion, this focus on ''words',' as a measure of language learning progress, is somewhat misleading. It leads me to think that the course authors made a decision early on to focus on learning vocabulary and designed the course around that. It's a fairly easy metric to get at and track. People tend to equate number of words known to fluency. Grammar and usage, on the other hand, much more subtle.

Wish me luck! I'm off to translate ''It's a tree.'' ''I am at the store''!


Hi again

  • Finally, the selection algorithm that chooses which words are presented in an exercise doesn't seem to chose the weakest words; instead, it appears to choose semi-randomly from among the weakest ~30 words [...]

That's what annoys and should be changed. I've reviewed a lesson over 10 times getting the same words [and I earned 10 lingots, if that says something...]. So the reviewing algorithm is actually not wrong, but it is the "weak words choosing" one.

.-. . .- -.. / - .... .. ... / .- -. -.. / - . .-.. .-.. / -- . / .-- .... .- - / -.. --- . ... / .. - / ... .- -.--


Thanks for the thourough explanitation!


If you just rush over it and never care about It for the rest for your life then what is the point of it? I remembers how many times you have wrote or said this word and how many times you got it wrong. it helps you remember the word.

that is exactly why I love duolingo.


No one would disagree with that. Bear in mind that this thread is nearly 2 years old (November of 2014). No one is complaining that words are reviewed at all, but rather how and when words were presented for revision by Duolingo using the "Strengthen Skills" option at that time.

I haven't regularly used Duolingo for over a year, but lately I've been using it again a bit as I'm encouraging my daughter to get started with it. Since I've been using it again, mostly on the app, I've noticed that the entire discussion here seems to have been rendered irrelevant by changes that Duolingo has made to the system sometime in the last 22 months. The "Strengthen Skills" option now seems to work as it should, focusing on specific weakened skills rather than on whichever individual words needed strengthening.


I don't want to spam the boards with complaints, so I'll just vent on here. Today 5 lessons lost gold, and it is the fourth day in a row where V. Inf 2 has lost gold. I gilded Medical yesterday, it has lost gold today. This can't be normal behavior.


Actually, this is exactly what I'd expect, and it's consistent with my earlier explanations. With larger skills come a larger number of words (e.g. 69 word forms in Inf.2 and far more than that in Medical - 59 lexemes and probably close to that in additional forms when plurals and gender forms are factored in). By practicing only until the skill regilds, you're practicing only until the average word strength for that skill just reaches 80% of its max. Meanwhile, other words that you haven't revised will continue to degrade, and that skill's average will drop back below 80 in short order.

Skills tend to degrade in clusters like this because when initially learning a skill, you generally complete its lessons within a few days of each other, depending on what sort of progress you make while learning. As a result, they tend to get reviewed around the same time and so they'll stay clustered together.


For the larger lessons it works much better for me if I redo the entire lesson rather than only the "strengthen" lesson. At least that gives me several days before it decays again. The algorithm is not very accurate however. If it focused on the words/lessons that the user habitually gets wrong, it would be far more useful, and i would be inundated with lessons on Habría vs. Hubiera.


Well, of course. If you repeat the entire skill, you're reviewing and strengthening every word that the skill teaches. That's basically the same result you'd get from repeating the skill-specific strengthening exercises an equivalent number of times. Actually, you'd need to do fewer strengthening exercises to achieve the same result (provided they're skill-specific) because each strengthen exercise targets 9 words for improvement, while lessons by default only target 7 words per lesson. For example, if you have a 10 lesson skill group with a total of 72 words, 8 practices (8*9==72) to cover the same number of words that repeating the 10 lessons would cover.


This is very useful thanks. Yes, whichever way I redo a lesson, I tend to redo it multiple times, because invariably there will be words and constructions than do not appear during the 1st several tries. Interesting to learn that some of them never appear.


I see what you mean. I guess I just question the value of this method, because more often than not I just get disheartened than anything. With German or Irish, I don't mind going back and regilding because I still feel like I am making progress. Just venting.


I understand, but bear in mind that this isn't really designed to help those of us that might already have a fair degree of familiarity with a language. It's meant to accommodate users who are learning the language for the first time that will require a lot of reviewing to lock the vocabulary into their long term memory. We can use it to refresh our skills, but we have to remember that this isn't what Duo is optimized for.

As an example, I learned a fair bit of German when I was stationed there while in the Army during the early 90's. I also lived in there for a while as a civilian once I was out of the Army. I used Duo to refresh my rusty skills, and because I went through it relatively quickly and don't feel the need to constantly practice it, my tree is nearly impossible to keep gold (for the same reasons you're experiencing with Spanish). This is true to an extent for French as well - I took it in high school and college and am not really trying to learn it from scratch here. On the other hand, I've practiced Spanish a lot in my time on Duo because I've learned it, more or less, from scratch. It stays gold with because I've built a history of lots of successes without peeking and without a lot of wrong answers. As a consequence, the decay rates are more stable and manageable.

Ultimately, I had to allow myself to be comfortable with the fact that it just doesn't matter whether my German or French trees are golden as long as I'm comfortable with what I know.


Does the number of times a user peeks at the hints affect the decay rate? If so, I'm encouraged by the fact that my peeking rate is diminishing as my familiarity with the language increases.


Thank you for this. It is something I'll just have to learn to accept too!


I've also had issues with strengthening skills. Sometimes I see a lesson which is 4/5 so one pass through should get it to gold right? Sometimes it doesn't even when I get everything right in the lesson. Very strange.


There's nothing strange about it, really. The strength of a skill is the average strength of the words taught within that skill, and that average has to be elevated above 80% of full strength for the skill to become golden. A skill with 10 lessons, could have more than 70 words (and even more word forms when gender, plurals or verb conjugations are considered). A practice session targets 9 words, or fewer if you do not complete a timed session. For most mid to large-sized skills, one skill-specific practice session won't be sufficient to bring the average up enough for it to become gold. See my post above for more details, specifically the two linked posts.


Ok, I see what you're saying and it makes sense. Thanks for explaining that to me!


Thank you, pinkodoug. It is great to know how this works. So, duolingo counts words practiced, not concepts (which would probably be hard to count). I am curious how you know these details. Are you or have you been an employee of duolingo?


Nope, I'm not an employee, nor have I been. Most of this is based upon careful observation of the app's behavior, examining the data sent to and from Duo, measuring results on words/vocabulary page (e.g. specific word strengths are buried within this data, and decay rates can be estimated) as well as from information stated by various devs over the last couple of years. Duolingo exposes quite a lot of information to the client, even if it's not directly presented on the page. If you know how to get at this information, the picture becomes clearer. For example, inspecting the data sent to the client when a practice session is loaded reveals, among other things, which words are being targeted for improvement. From there, it's not hard to deduce what's happening on the back end based on observable behavior over a number of sessions.


Wow, it is clear you have put a lot of time into figuring all this out! Sure appreciate your effort here.


If I go to the Words page and sort them by Strength (lowest at the top), I can see all the really decayed words. If I then use "Review flashcards" (from the right side of that same page), I think it is working on the most decayed words, and I do see the number of really decayed words decrease.

I wish they'd show a breakdown of word counts by decay level. I estimate the count by how many times I have to click Page Down to get to the next decay grouping, on that page.

Learn Spanish in just 5 minutes a day. For free.