coffee's guide about

Strategies to avoid memory interference in flashcard collections
 ( how to stop mixing up anki reviews )

What is interference?

Sometimes new or old flashcards make us confuse the answers because they create some intuition or pattern that is erroneous. For this there are many strategies we can try.

How to deal with multiple possible answers?

Add a hint to the front to disambiguate

Example 1:

q: 明日 (formal)

a:   あす

q: 明日

a: あした

Example 2:

q:  今日 (common)

a: きょう

q:今日 (rare)

a: こんにち

Example 3:

q: 脅かす(common)

a: おどかす

q: 脅かす(rare)

a: おびやかす

Using the same hint for the same kinds of ambiguities is a good idea to avoid relying on the hint too much.

In the case of using rarity as a hint consider adding some kind of simplified rating to the back of the card so that you get a better idea of what rare means in the specific context. a 5 star system might be better than a specific percentage. The last thing you want to be doing is memorizing numbers.

I keep confusing two flashcards, What do I do?

Make a disambiguation flashcard. You can make a separate deck for disambiguations or a tag so that you can create a filtered deck to cram once in a while if you feel like it. A disambiguation flashcard can look like this:

q: which context makes 明日 have two readings?

a: formality

A disambiguation card might not give a full answer for the fact in question and instead show the facet that specifically disambiguates the answers. This can avoid making the actual reviews of the cards easier than they should.

Tagging disambiguation cards can help with cramming sessions if you feel that the interference is really problematic at some point in the future. These cards are very easy to review since they only test the disambiguation aspect.

Another example:

q: 明日 (あした) alternative reading?

a: あす(formal)

Writing the element you are the most familiar with or the answer you default to as part of the question makes it so you are using that one as an anchor for the exception or hard to remember fact. This is particularly useful when one of the items is so familiar that learning a new fact about it is hard.

Disambiguation for a list of facts related to the same question:

Problematic example:

q: quadrilateral

a: square, rectangle, trapezoid

Proposed solution:

q0: what is a quadrilateral?

a0: a geometric shape that has four sides

q1: quadrilateral examples (..., rectangle, trapezoid)

a1: square

q2: quadrilateral examples(square, ..., trapezoid)

a2: rectangle

q3: quadrilateral examples (square, rectangle, …)

a3: trapezoid

q4: quadrilateral examples (all)?

q4: square, rectangle, trapezoid

Exceptions to patterns:

Sometimes patterns emerge in groups of facts that create problematic intuitions. We can try to exploit the awareness of the pattern to learn the exceptions.

Pitch accent deck:

As a motivating case, here's a list of words that formed a pattern for the learner:

蒼井さん、新井さん、石井さん、今井さん、藤井さん、松井さん are all heiban.

桜井さん、長井さん are accented but not in the same way.

This created a bias towards heiban.

Example of naive cards:

q: 蒼井さん

a: heiban

q: 新井さん

a: heiban

q: 石井さん

a: heiban

q: 今井さん

a: heiban

q: 藤井さん

a: heiban

q: 松井さん

a: heiban

q: 桜井さん

a: さくらいさん(2)

q: 長井さん

a: ながいさん(1)

The problem is with the last two answers. All the other ◯井さん reinforce the same answer.

We can't use context, order, or a rule. There might be a rule but for the purpose of this example I assume that this is arbitrary.

First strategy: odd one out

Example cards:

q: 藤井さん    松井さん    桜井さん

a: heiban



We can attempt learning which one is different even before we learn the specific pitch for the odd one. Some scripting might help with randomizing the order in which the words appear to avoid “learning the card” instead of learning the fact. Alternatively making more cards that are identical but with different order can be achieved with an anki template like the following one:

card 1:




card 2:




card 3:




card 4:

Word1, Word2, and Word3 are fields in the anki note.You’d need as many Answer1, Answer2 and so on to match on the back.

This kind of template would generate all the shuffling you need and work on anki mobile.

Second strategy: exception focus

We can first try to add to the front that the word is an exception and only test the pitch:

q: 桜井さん (exception)

a: さくらいさん(2)

q: 長井さん (exception)

a: ながいさん(1)

Or we can test for remembering which are the exceptions:

q: Exceptions to 蒼井、新井、石井、今井、藤井、松井 ?

a:  桜井さん : さくらいさん(2)

     長井さん : ながいさん(1)

Combined with the group strategy:

q: Exceptions to 蒼井、新井、石井、今井、藤井、松井 ? ( 桜井さん, … )

a:  桜井さん : さくらいさん(2)

     長井さん : ながいさん(1)

q: Exceptions to 蒼井、新井、石井、今井、藤井、松井 ? ( … , 長井さん )

a:  桜井さん : さくらいさん(2)

     長井さん : ながいさん(1)

50-50 situations:

Even though sometimes answers can feel 50-50 undoubtedly we’ll favor one more than the other.

Since we can’t focus on context, rules or order, we need to exploit either some specific feature of the words, as an ad hoc rule, or a mnemonic. Ad hoc rules can be easier to make if you already master most of the material; for example stroke count of the kanji involved, a group the problematic kanji can be grouped in (like living things, sharing a radical, rarity, etc.), readings in another context or any other element that is already known. This is the most time consuming method but if some elements of the group are already solid in your memory the process is much easier.

Example of learning a grouping:

q: Why are these a group?


a: heiban

q: Why are these a group?


a: exceptions to the heiban pattern in  蒼井、新井、石井、今井、藤井、松井

q:  Why are these a group?


a: concrete nouns. they are part of the heiban pattern 蒼井、新井、石井、今井、藤井、松井 and the last one is an exception 桜井さん : さくらいさん(2)

q: Why are these a group?


a: abstract nouns. they are part of the heiban pattern 蒼井、新井、石井、今井、藤井、松井 and the last one is an exception 長井さん : ながいさん(1)

Concreteness of nouns is subjective, so keep that in mind. Make groups that make sense to you.

Some audio approaches

Pairing words:

q: (audio of the problem word or sentence) (audio of another thing that you know the answer for)

a: answer that matches both

You might not even need an answer, the joint association might be enough.

Contrasting words:

q: (audio of problem word)(audio of known word) Difference?

a: 1 is heiban, 2 nakadaka.

The odd one out:

q: (audio of a bunch of words with an exception at the end or at the end) Odd one out?

a: explanation of the pattern or exception

There are some prejudiced views against "audio in front" flashcards because you might "learn the card and not the information".  In this case, when it comes to resolving confusion, just getting tired of the card means that you improved the skill of feeling the difference.

Randomization or shuffling as discussed in the template example is a good idea too, although it can be too much given how long audio cards can take to review compared to plain text.

Closing words

There are many strategies available beyond rote repetition and you can take full advantage of without much additional effort. Try to see what works and remember that interference is something you can exploit for your own gain.

Try these strategies and let me know what works for you.

Share this page with people who might have encountered similar issues.

Main Site