53

The Forgetting Curve and Learning Algorithms

Jeff Hanks!and Ping Zhan!!

Abstract

With the globalization and impact of Internet, a new wave of English as public language is
occurring in Japanese corporations. Learning English in an efficient ways is still a challenge. In this
manuscript, we introduce the forgetting curve and spacing effect, algorithms in the software
SuperMemo. We also propose an experiment project on how to test learning English words based
on the spacing effect. A pre-experiment result is also given.

Index terms: forgetting curve, spacing effect, spaced repetition, SuperMemo, learning algorithm,
experiment.

1. Introduction

There are now (1,750,000,000) one billion, seven hundred and fifty million people speaking
English globally. In Europe and Asia, the non-English-speaking world, the number of companies
where English is used as an official language is growing. Rakuten−the Japanese biggest online-
shopping, announced this change recently and brought a lot of attention. It has been said that
newly hired employees in Samsung Electronics, the South Korean multinational electronics
company, are required to score over nine hundred on the TOIEC test. Other companies, such as
Airbus, DaimlerChrysler, Nokia, SAP, Microsoft in Beijing and Tokyo, also use English the official
language in the company [4].
The reason that a language strategy is needed now is that global business speaks English.
Besides this, there are three more factors to be considered: The gradual downsizing of the Japa-
nese market, the effect of the Internet and English media, and finally the rapid Englishnization
in Asia. It is inefficient to communicate in multiple languages and English becomes a must in
order to compete globally. To achieve Englishnization, there are still a lot of hurdles. The big-
gest one is the English level of individual employees [4].
One author is now researching on the topic in relation to supply chain management, where
the phenomenon in that fluctuation in orders increase as one moves up the supply chain from
retailers to wholesalers to manufacturers to suppliers is referred as the bullwhip effect. The
bullwhip effect results in an increase in all costs in the supply chain and a decrease in customer
service levels. Therefore, one objective of coordination is reducing the bullwhip effect. A lot of

2012 年 11 月 30 日受付
!江戸川大学 情報文化学科専任講師 語学
!!江戸川大学 情報文化学科教授 数理計画
 54 The Forgetting Curve and Learning Algorithms

managerial levels have been discussed. To make these managerial levels work, trust between
partners is important and this trust can only be based on regular and relatively long-term com-
munication, in the global case, a multicultural communication [6].
Now with the high-speed development, we have to learn more and learn more efficiently.
Learning English based on the mechanism of the brain is the motivation for our work. A re-
searcher of science and medicine, Professor Kuniyoshi Sakai researched the issue.

“English or Japanese, although they are two different languages, the same parts of brain are
used. . . .The most important thing is whether it is a first language, which is learned before
7 years old, or a second language, which is one learned after 7 years old. . . .There are four
separate parts of brain used in language processing, words, intonation or accent, grammar,
meaning. If the field responsible for grammar has been damaged, it leads to language disabil-
ity, which has been clarified recently. . . .The process of learning language: When nothing
about the knowledge of grammar changes, the activity of brain does not change. Then the
activity of brain is in proportion to the knowledge of grammar. After six years or so, with
high level of grammar, a lot of examples, conversely, show an energy-savings taking place in
the brain. . . .Although, like other fields, there are some disputes, that is, two assumptions:
leaning−stimulated from their parents (古典な後天説), and acquisition−developed naturally
with growth, an inborn ability (近年の先天説).” [5].

This is beyond our scope now. However, it is clear that there are some rules about learning
language regarding the mechanisms of the brain. Knowing these rules and finding some efficient
ways to learn English or other languages is the motivation of our current research.
The manuscript is organized as follows. We begin by reviewing the forgetting curve and
spacing effect in Section 2. In Section 3 we introduce the algorithm in the software SuperMemo.
A pre-experiment outline is discussed in Section 4 and a related simple experiment result is
shown in the final section.

2. Forgetting curve and spacing effect

The famous forgetting curve was hypothesized by Hermann Ebbinghbaus in 1885. His curve
proposes to show how memory of a piece of data declines over time when there is no attempt to
reinforce it. As items are reinforced the curve flattens, until the data is retained long-term.

“Ebbinghaus. . .used himself as a subject and memorized lists of nonsense syllables until they
could be perfectly recited. Later after varying delays of up to 31 days, he relearned those
same lists and measured how much less time was needed to learn them again relative to the
time required to learn them in the first place. If 10 minutes were needed to learn them again
after a delay of 6 hours, but only 4 minutes were needed to learn them again after a delay of
6 hours, then his memory was such that 60％ savings had been achieved. As the retention
interval increased, savings decreased, which is to say that forgetting occurred with the
 The Forgetting Curve and Learning Algorithms 55

Source: Herman Ebbinghaus, Memory, a Contribution to Experimental Psychology, 1885.

Figure 2!1 Ebbinghaus’ classic forgetting curve.

passage of time.” [8]

By charting this decay over a period of 31 days produced his now famous forgetting curve
(Figure 2!1). Ideally, reinforcement should come just before an item is forgotten, but not too
soon. This is termed the “spacing effect”. The spacing effect shows that reinforcement over a
longer period is more effective than repetition over a short period of time (cramming). This is
where “spaced repetition” comes in.
Spaced repetition is a method of learning designed to take advantage of the spacing effect.
It was first proposed by C.A. Mace in Psychology of Study in 1932. Pimsleur language courses,
developed in the 1960s, are the most famous example of second language teaching based on
spaced repetition. The Leitner system, developed by Sebastian Leitner in the 1970s is a method
of using flashcards based on spaced repetition.
Mathematical models of forgetting have also been researched in industrial production. Three
types of models are classified, VRVF (Variable Regression to Variable Forgetting), VRIF (Vari-
able Regression to Invariant Forgetting) and LFCM (Learn-Forget Curve Model) [3] These
models were developed in an attempt to maximize worker production by determining the opti-
mum length of breaks for factory employees in order to minimize the forgetting of tasks upon
returning to their work. As these models focus on tasks as opposed to words or data, they are of
little use to us in our proposed experiments.
 56 The Forgetting Curve and Learning Algorithms

3. Learning algorithm

Learning algorithms about some special patterns have been developed in machine learning,
artificial intelligence and neural networks. Surprisingly however, there is little research into the
optimal spacing of repetition for new ESL vocabulary. To our knowledge, referenced papers on
a general learning algorithm have not been published in official academic journals. Here we
introduce the learning algorithm in the software SuperMemo provided on its web site.
SuperMemo is the learning software that we intend to use in our work. SuperMemo, developed
in Poland by Piotr Wozniak in 1985, was the first example of PC software created to exploit
spaced repetition. Since that time, a number of programs and web based applications have been
developed, one of the most popular of which is “Anki”, an open source flashcard application
which is available on multiple platforms, including most recently, smart phones. For ease of
understanding, we begin with the early version of the algorithms, Version 2 [7].

First, we define some notations.

I (n) : inter-repetition interval in days,

EF : easiness factor reflecting the ease of memorizing and retaining a given item in memory,

q : quality of the response, the 0!5 grade scale defined as follows:

5 !correct immediate response,
4 !correct response with hesitation,
3 !correct response recalled with serious difficulty,
2 !incorrect response; where the correct one seemed easy to recall,
1 !incorrect response; the correct one remembered,
0 !complete forgetting.

Initializing :
EF : ＝ 2.5,
I (1) : ＝1,
I (2) : ＝6.

The main procedure, the recursive formula :

I (n) : ＝I (n-1)・EF,
EF’ : ＝f (EF,q)
for all n＞2.

Where EF is restricted between 1.1 and 2.5, performs better if EF is larger than 1.3 (if EF is less
than 1.3, set it to 1.3) and f (EF, q) is given as EF＋(0.1−(5−q)・(0.08＋(5−q)・0.02)).
 The Forgetting Curve and Learning Algorithms 57

Termination:
If q＝4 or q＝5.

Re-initializing:
If q!2, set I (1) : ＝1, I (2) : ＝6, and without changing EF.
(q)＝(0.1−(5−q)・(0.08＋(5−q)・0.02)), note the values of f!
Denote f! (q) are −0.8, −0.54,
−0.32, −0.14, 0, 0.1 for q＝0,1, . . . , 5. And if q takes a lower value, the interval for repetitions is
shorter. Since recursive occurs when q"4, EF is decreasing. Also keeping q larger than 1.3 means
the interval is increasing.

In version 4 of the algorithm, there are two steps in the recursive formula

Step 1 : temp＝I (n−1)＋I (n−1)・(1−1/EF)/2・(0.25・q−1)

Step 2: (1−fraction)・I (n−1)＋fraction・temp.

Where temp is an auxiliary value used in calculations, fraction is any number between 0 and 1.
Note for q"4, temp"I (n−1, EF) and also the result of Step 2, which is inconsistent with the
concept of spacing effect. And it is revised in version 5.

In version 5, a new factor, optimal interval factor OF is introduced. And the recursive steps are
as follows:

I (n, EF)＝OF (n, EF)・I (n−1),

where OF is calculated in two steps.

Step 1 : OF・(0.72＋q・0.07)
Step 2 : (1−fraction)・OF＋fraction・OF!

Note again that 0.72＋q・0.07"1 if q"4.

Although EF as a function of q is renewed with the same formula f, no details about the
relation of EF and OF have been shown. This is also the case in later versions of algorithms. In
Version 11, Not OF but dOF, the decrement of OF is recursively calculated, and EF, easy factor,
is replaced by AF, Absolute difficulty Factor, the description is also not complete.

4. Pseudo Optimal interval proposal

The ultimate objective is to find the pseudo-optimal inter-repetition interval. One opinion
holds that the optimum inter-repetition interval is likely to be the longest interval that avoids
retrieval failures. These are the maximum intervals that the human memory is capable of
 58 The Forgetting Curve and Learning Algorithms

between forgetting and remembering. This is consistent with the spacing effect and also coin-
cides with the following in some sense. In chance theory or credibilistic theory (可信性理論) or
uncertainty theory, 0.5 (the probability value, or the fuzzy characteristic value) is the most
noticed value, in this case (farthest distance between the success (remembering) or failure (for-
getting). At this point brains work most actively.
Our pre-test procedures are just for an intuition into understanding the direction of the
experiments. What is possible and what are the problems we need to solve? With so many
variations of students’ backgrounds and variations in words, there is a question as to whether we
can obtain results that verify our assumptions.
In Section 3, we see that with the fact of easiness, the intervals are vary from 1.1 to 2.5, that
is, vary from the unit interval and double roughly the former interval. For the [double], it is not
only the medium, also with the accumulation of the words in previous days and some repetition,
we HAVE TO and CAN decrease the time necessary.

The simplified version for pre-testing experiments is as follows.

1) Based on Ebbinghaus’ famous forgetting curve, that is, one hour, the next day, the week-
end, and the end of month.
2) Double the former interval (medium, or average procedures), Note that the intervals here
also form a geometric sequence with the factor 2.
3) If the experiments are done in the class lecture, a geometric sequence may be difficult
because it would last more than one semester, in this case, we approximate it with a
sequence, a−1, a−2, . . . , a−n, . . . , with a−(n＋1)−a−n＝n.
4) Equal intervals (a common or simple one for comparing)

An approximate example of a geometric sequence appears at http://www.supermemo.com/

help/faq/algorithm.htm.

1) The 1 st repetition should be scheduled 30 minutes after reading words for the first time
2) after an hour
3) after 9 hours
4) after 24 hours
5) after 3 days
6) after 6 days
7) after 12 days

Our experiments will concentrate on words because a vital component of language learning
is vocabulary. However in teaching a second language two questions arise. First is what vocabu-
lary to teach. The second is how to deliver that vocabulary to students in an effective way.
There are no other methods except for simple accumulation. Without a rich and varied vocabu-
lary, you can only transmit simple messages. It is hard to communicate your personality and
 The Forgetting Curve and Learning Algorithms 59

make an impact on others, which is one main shortcoming of Japanese (or Japanese leaders) now.
The learning process is relatively simple, that is, not dependent too much on personal factors.
As for what to teach, there are a variety of word lists used in ESL. The most famous of which
is the British National Corpus (BNC). In Japan there is the JACET 8000. Another alternative is
the General Service List (GSL) developed by Michael West, in 1953. There are pros and cons to
each of these lists, but for the purpose of our proposed research, we have selected the Academic
Word List (AWL), developed by Averil Coxhead at Victoria University in Wellington. It was
developed as a companion to the GSL, but focuses on words occurring frequently in academic
texts. The reason for selecting this list is that a majority of the words will be unfamiliar to our
experiment’s subjects (Edogawa University students), but will be potentially useful to them in
their careers.

5. Experiments

We are very fortunate that before we begin this work, one student of the Department of
Communication and Business, Tetsuya Inaba, who is interested in language and the mechanisms
of the brain, did some experiments on retention of English and German words according Ebbin
ghaus’ forgetting curve. We summarize some of them.
The words are from textbooks, 1,550 English words and 1,320 German words were selected.
The experiment period lasted from August to November. Inaba thought that German is much
more difficult to remember than English. He grouped 26 English words and 20 German words
into one unit. It was also assumed that the forgetting rate of German words would be much
higher than for English words. Table 5-1 is the result.
Where, the first column indicates interval Table 5!1 The remembering rate
time, the second and third columns indicate
Intervals English German
the average words remembered versus words
20 min 21 : 5 ｜ 26 18 : 2 ｜ 20
forgotten and the remembering rate.
Because the number of German words is 1 hour 23 : 3 ｜ 26 15 : 5 ｜ 20

only 20, less than the 26 of English words, 1 day 19 : 7 ｜ 26 14 : 6 ｜ 20

German had an advantage over English here. a week 19 : 7 ｜ 26 14 : 6 ｜ 20

The average remembering rate, 78％ of Eng- a month 19 : 7 ｜ 26 12 : 8 ｜ 20
lish, is unexpectedly near to 73％, the rating of
German. Mr. Inaba also said that he was a little astonished by the results.

Acknowledgments : The authors are very grateful to Tetsuya Inaba for providing us the pre-experiment
results.

References
[ 1 ] John Henderson, Memory and Forgetting, Routledge, 1999.
[ 2 ] Michael Jacob Kahana, Foundations of Human Memory, Oxford University, 2012.
[ 3 ] David A. Nembhard, Napassavong Osothsilp, An Empirical Comparision of Forgetting Models, IEEE
60 The Forgetting Curve and Learning Algorithms

Transactions on engineering management, Vol. 48, No. 3, 2001.

[ 4 ] 特集, グローバル英語力, Harvard Business Review, p. 108!115, 21012/10.
[ 5 ] 酒井邦嘉, Language Acquisition and Brain Development, (脳科学が明かす言語習得のメカニズム),
Harvard Business Review, p. 108!115, 21012/10.
[ 6 ] Sunil Chopra, Peter Meindl, Supply Chain Management 3rd, Prentice Hall, 2006.
[ 7 ] www.supermemo.com/english/ol/sm4.htm, http://www.supermemo.com/english/ol/sm2.htm
http://www.supermemo.com/english/algsm11.htm#Advanced repetitions
[ 8 ] Sergio Della Sala, Forgetting, Psychology Press, 2010.