Test of a Model for Predicting
Second Language Lexical Growth
through Reading
Marlise Horst
Concordia University, Montreal
What’s it going to be then, eh?
There was me, that is Alex, and my three droogs, that is
Pete, Georgie and Dim, Dim being really dim, and we sat
in the Korova Milkbar making up our rassoodocks what to
do with the evening, a flip dark chill winter bastard though
dry. The Korova Milkbar was a milk-plus mesto, and you
may, O my brothers, have forgotten what these mestos
were like, things changing so skorry these days and
everybody very quick to forget, newspapers not being read
much neither. Well, what they sold there was milk plus
something else. They had no license for selling liquor, but
there was no law yet against prodding some of the new
vesches....
(from A Clockwork Orange by Antony Burgess)
BACKGROUND
1. Has research shown that L2 readers learn new words
from reading?
2. Why are word learning gains so small in these read-andtest studies?
Text
Saragi
et al
1978
Ferris
C’work
Orange
Animal
Farm
Pitts
1988
1989
exp 2
Day
Hul- Dupuy & Horst
et al stijn Krashen et al
1991 1992 1993
1998
exp 1
C’work simp. simp. Trois
Orange text text hommes
simp.
text
Text
21,000
length
60,000
???
6700
Eng.
Dutch
1032
907
Eng.
video+
15 pages
No.of
45
items
tested
90
50
Mean no.
5
of words
learned
68
7*
28
2
17
3*
12
1
*gain established by comparison to a control group
7*
30
BACKGROUND
3. What do these studies leave unanswered?
• How well are words learned and retained?
• What happens to partially learned words?
• What happens when partially learned words are
encountered again?
• How well does reading work for individuals?
What can we expect?
BACKGROUND
4. What might a predictive model offer?
• something to test
• insights into the word learning process
• an eventual explanation?
• a way of quantifying the effectiveness of
reading
• a basis for making teaching and learning
decisions
RESEARCH QUESTIONS
1. How many new words did the subject learn?
2. How well were the words learned?
3. How well did the matrix model succeed in predicting
gains (and losses)?
METHOD
• case study
• 8 readings of same text
• adult learner of Dutch
• comic-book text: Lucky Luke:Tenderfoot
• tested on 300 words that occurred once (pre-test)
• read text weekly (Saturdays)
• tested on 300 words weekly (Wednesdays)
• delayed post-test 10 weeks later
TEST
0 = I definitely don't know what this word means
1 = I am not really sure what this word means
2 = I think I know what this word means
3 = I definitely know what this word means
NOTE HOW INCREMENTAL KNOWLEDGE APPEARS
IN MATRIX
After 1 reading
0
1
2
3
0
75
4
2
0
1
27
20
4
0
2
9
20
13
7
3
3
6
35
75
2
6
20
12
10
3
2
6
32
109
After second reading
0
1
2
3
0
53
4
2
0
1
19
20
4
0
Notice:
The diagonal represents words that have
not changed position.
Words below the diagonal are losing
ground; above are gaining.
With each reading, words are going in
both directions.
But, over readings, more words are
heading up than down!
RESULTS
1. How many new words did the subject learn?
30 0
28 0
26 0
24 0
22 0
20 0
18 0
W
O 16 0
R
D 14 0
S
12 0
10 0
80
60
40
20
0
Pre-test
1s t Read ing 2n d
SURE I KNOW
3rd
4th
THINK I KNOW
5th
NOT SURE
Rated 3 on pre-test:
82
Rated 3 after 8th reading 223
Gain:
141
Gain after 1st reading:
Gain after 2nd reading:
37
45
6th
7th
8th
DON'T KNOW
de layed post
RESULTS
2. How well were the words learned?
30 0
28 0
26 0
24 0
22 0
20 0
18 0
W
O 16 0
R
D 14 0
S
12 0
10 0
80
60
40
20
0
Pre-test
1s t Read ing 2n d
SURE I KNOW
3rd
THINK I KNOW
Rated 3 after 8th reading:
Rated 3 delayed post-test:
Loss:
Rated 3 on pre-test:
4th
5th
NOT SURE
6th
7th
8th
de layed post
DON'T KNOW
223
198
25
82
90% of words rated “definitely known” (3) translated
correctly
RESULTS (MATRIX PROCEDURE)
3. How well did the matrix model predict gains?
RAW DATA MATRIX
Time 1 (after 1st reading)
state
0
1
0
75
2
27
3
9
3
=114
Time 0
1
(before
reading)
4
20
20
6
2
2
4
13
35
3
0
0
7
75
How to read this data:
After one reading, some of the 114 words originally rated 0
(unknown, 75+27+9+3) had moved to other categories
75 had remained in category 0
The probability that a word rated 0 at pretest was still rated
0 after one reading was 75/114 = .66
… that a word rated 0 had moved to 1 (not sure) was
27/114 = .24
Same calculation for all cells  probability matrix
PROBABILITY MATRIX
Time 1 (after 1st reading)
state
0
1
2
3
0
.66
.24
.08
.02
1
.08
.40
.40
.12
2
.04
.07
.24
.65
3
.00
.00
.09
.91
Time 0
(before
reading)
The probability, multiplied by the original number, equals
the number in the raw data above.
So 114 original zero ratings, x .66 probability it will remain
zero, = 75 words remaining at zero after one reading.
And so on for all cells
Now we can use the probabilities to predict how many
words will remain at 0 after a second reading
75 x .66 = 49.5
And a third
49.5 x .66 = 32.67
And so on, for all cells, through any number of iterations,
making predictions that can be compared to human data.
RESULTS
3. How well did the matrix model succeed in predicting
gains?
REPORTED WORD KNOWLEDGE
No. of words at state:
0
1
2
3
114
50
54
82
1
81
51
49
119
2
72
27
37
164
3
57
33
37
173
4
48
30
41
181
5
40
30
39
191
6
43
26
31
200
7
39
24
28
209
8
30
20
27
223
pre-test:
reading
round:
MATRIX PREDICTIONS
No. of words at state:
0
1
2
3
2
60
43
49
3
45
35
47
173
4
34
28
45
193
reading
round
147
5
27
23
42
209
6
21
18
40
220
7
17
15
39
229
8
14
13
37
236
6
7
34
253
16+
if iterated….
RESULTS
3. How well did the matrix model succeed in predicting
gains?
Predictions and performance: "definitely known" and “think I know” ratings
Based on multiplier from first reading, then iterated for each additional.
Figure 8.3
Predictions and performance: "not really sure" and "definitely don't know" ratings
CONCLUSIONS
• An accurate and useful predictor?
• Predicts incidental acquisition?
• A lot of word learning occurred — some complete, some
partial.
• Eight reading encounters leads to learning that sticks.
• Reading a text twice makes a big difference.
• Give language learners comic books to read?