Are justification and hyphenation good or bad for the reader? Leyla Akhmadeeva Rinat Gizatullin Boris Veytsman Bashkir State Medical Univeristy, Ufa, Russia George Mason University, USA TUG’2016
Introduction We all used to consider justified text “better”. Compare: As any dedicated reader can clearly see, the As any dedicated reader can clearly see, Ideal of practical reason is a representation the Ideal of practical reason is a of, as far as I know, the things in them- representation of, as far as I know, the selves; as I have shown elsewhere, the phe- things in themselves; as I have shown nomena should only be used as a canon for elsewhere, the phenomena should only be our understanding. The paralogisms of prac- used as a canon for our understanding. The tical reason are what first give rise to the ar- paralogisms of practical reason are what chitectonic of practical reason. As will easily first give rise to the architectonic of be shown in the next section, reason would practical reason. As will easily be shown in thereby be made to contradict, in view of the next section, reason would thereby be these considerations, the Ideal of practical made to contradict, in view of these reason, yet the manifold depends on the considerations, the Ideal of practical phenomena. Necessity depends on, when reason, yet the manifold depends on the thus treated as the practical employment of phenomena. Necessity depends on, when the never-ending regress in the series of em- thus treated as the practical employment pirical conditions, time. Human reason de- of the never-ending regress in the series of pends on our sense perceptions, by means empirical conditions, time. Human reason of analytic unity. There can be no doubt that depends on our sense perceptions, by the objects in space and time are what first means of analytic unity. There can be no give rise to human reason. doubt that the objects in space and time are what first give rise to human reason. But is it really better for reading and comprehension? 2
Experiment setup N = 300 healthy volunteers. T wo texts, A and B . ParaT ype Cyrillic, standard L A T EX setup. Half volunteers get A justified, B ragged. Another half gets A ragged, B justified. 1. Each volunteer reads text A , and then text B . 2. The speed of reading is measured. 3. Volunteers answer 10 multiple choice questions about each text (immediate test). 4. After 60 min the test is taken again (delayed test). Comparisons: speed of reading, immediate test, delayed test. 3
Direct comparisons and t -test Speed of reading, J − R 200 ● ● Words per min ● ● ● ● ● ● ● ● ● ● ● ● 50 −100 ● ● ● ● ● ● ● ● ● ● ● ● No difference ( p = 0 . 12) 4
Immediate test, J − R Correct answers 2 0 −4 ● No difference ( p = 0 . 17) 5
Delayed test, J − R Correct answers ● ● ● ● ● ● ● ● 2 ● ● ● ● ● ● ● ● ● ● 0 ● ● ● ● ● ● ● −4 ● ● ● ● ● ● ● Ragged is better ( p = 0 . 001 ). Mean − 0 . 26 , standard deviation 1 . 36. 6
Problems with t -test Individual differences are large. Differences between texts’ diffi- culty level are smaller but essential . T ypographic differences are small. We use paired tests, but 1000000000000000000 − 1000000000000000000 + . 00001 = ? 7
Problems with t -test Individual differences are large. Differences between texts’ diffi- culty level are smaller but essential . T ypographic differences are small. We use paired tests, but 1000000000000000000 − 1000000000000000000 + . 00001 = ? Bayesian statistics: Bayesian modeling: 1. Construct a model of the effect 2. Start with broad assumption of the parameters 3. Using Bayesian theorem get updated values of model parameters. 7
Bayesian statistics Model Speed of reading: a linear model V = V individual + (Correction if text B ) + (Correction if ragged) 8
Bayesian statistics Model Speed of reading: a linear model V = V individual + (Correction if text B ) + (Correction if ragged) Cannot do the same with probability to answer a question p : we can get p > 1 or p < 0. 8
Bayesian statistics Model Speed of reading: a linear model V = V individual + (Correction if text B ) + (Correction if ragged) Cannot do the same with probability to answer a question p : we can get p > 1 or p < 0. Let us introduce log odds: � p � L = l n 1 − p Linear model again L = L individual + (Correction if text B ) + (Correction if ragged) 8
Method Gaussian prior for the individual contribution and wide uniform priors for the other parameters. Multiple Chain Monte Carlo simulations (16 chains with 10,000 sam- ples each) for each model. 9
Results: speed of reading Individ. diff A−B J−R 0.12 0.12 0.10 0.08 0.08 0.08 0.06 0.04 0.04 0.04 0.02 95% 0.00 0.00 0.00 60 70 80 90 5 10 15 20 25 30 35 −5 0 5 10 15 20 Words per min Words per min Words per min Justified is faster: about 7 words per minute difference. 10
Results: immediate test Individ. diff A−B J−R 5 4 4 4 3 3 3 2 2 2 1 1 1 95% 0 0 0 0.6 0.7 0.8 0.9 1.0 1.1 1.2 −0.2 0.0 0.2 0.4 −0.4 −0.2 0.0 0.2 Log odds Log odds Difference J−R, log odds A suggestion that ragged is better. . . 11
Results: delayed test Individ. diff A−B J−R 5 4 4 4 3 3 3 2 2 2 1 1 1 95% 0 0 0 0.8 1.0 1.2 1.4 −0.2 0.0 0.2 0.4 −0.6 −0.4 −0.2 0.0 Log odds Log odds Difference J−R, log odds Ragged is definitely better. . . On a 100 questions test with 90% correct answers the difference would be about 4 points. 12
Discussion 1. What does this tell us about “whole word” reading? 2. Is the effect culture dependent? 3. What is the role of Cyrillic script? Does this work with Latin script? 4. Does the effect depend on the language proficiency? 13
Conclusions 1. We have a quite strange result: ragged is better for reading comprehension, but slightly worse for speed of reading. 2. We still do not quite understand the meaning and implications of this result. 14
Recommend
More recommend
