Aentlpoe
1
Contents 1 Visual Perception 4-5 2 Geometric Illusions 8 3 Graphical Perception 9 4 Word jumble hoax debunked 11 5 Information retrieval of jumbled words 11-19
6 Does the human mnid raed wrods as wlohe? 21
Baer
2
Buttrelfy
3
Visual Perception
The human eye-brain system is arguably the most sophisticated computing system which we have access to. It can easily handle complex visual processing and pattern recognition tasks which would be impossible to attempt on even the most powerful supercomputer. If we are going to use our visual skills to assist us in data analysis, it is important to remember that they have evolved to handle quite different tasks from those encountered in a typical data analysis. The tasks which our visual system excels at are those which were useful to our hunter-gatherer forbears. These tasks include recognising shapes, discerning colour, judging sizes and distances, and tracing and extrapolating motion in three dimensions. Some of these skills seem directly useful in data analysis, but it is very important that we understand both the strengths and weaknesses of the visual system when used in this way. Despite its awesome power, we tend to take our visual system for granted; perhaps because we make use of it in virtually every task we perform. One assumption that we make is that we can trust what we see. After all, “seeing is believing�. In fact this is not always the case, and sometimes we can fooled. This is revealed by the existence of visual illusions and we’ll look at a number of these now. Most of the time our eyes give us a good sense of the way the things are. Sometimes, however, we can be quite misled into seeing effects which are not really present. Images which produce this kind of phenomenon are called visual illusions. Figure 5.1 is a spectacular example of a visual illusion. It is hard to believe, but all the lines in this figure are either horizontal or vertical, and all the black and white polygons are squares. To check this, you should try lying a piece of paper along the horizontal lines. The illusion was first discovered by accident when workmen used the pattern to decorate the outside wall of a cafe with a black and white tiles. The existence of visual illusions indicates that we need to be careful when using graphics in data analysis. If an apparent feature in a graph is due to a visual illusion rather than a real effect, then we may draw the wrong conclusions from our analysis. There are many kinds of illusion, some are purely geometric, and others related to colour perception. In this section we will examine a few of the more famous geometric illusions and see what consequences theCaeml existence of these illusions may have for data analysis.
4
Perspective Illusions We saw in section sec:perspective that the human eye acts as a “pinhole camera” and that this results in the image of an object which appears on our retina as being larger when that object is close than when it is far away. Our brain however corrects our interpretation of the size of objects using any additional knowledge about how far away the object is. When we are mislead about how far an object is away, we can also be mislead about its size. The Ames room leads us to misjudge the relative size of objects it contains because it deliberately gives false cues about distance. We can also be lead to misjudge the size of things in two dimensional pictures because we use depth cues based on apparent perspective to “correct” the size of objects.
Crcoodlie
5
Ealge
6
Elpehnat
7
Geometric Illusions
shows such a perspective illusion. The perspective is emphasized because many of the lines in the picture converge to a common vanishing point. The “figures� in the picture are all the same size, but because the appear to be progressively further away, we see them as progressively larger. In this figure there are obvious features which suggest perspective. Sometimes the cues are more subtle, and we are not always aware that we are being mislead. The Ponzo Illusion Figure 5.2 has a clear interpretation as a three dimensional scene, so it is no surprise that perspective influences our perception of the size of objects in it. The effects of perspective an be much more subtle however. The Ponzo illusion (named after the Italian psychologist Mario Ponzo) is a famous example of how equal length lines can be perceived as having different lengths. The lower of the two horizontal line segments appears to be shorter than the upper one. One explanation for this is that the sloping lines create the same impression of depth as, for example, railway lines. A line of a certain length is perceived as being longer the further away you think it is. The phenomenon shows how the visual system tends to treat figures as three dimensional.
Elpehnat
8
Graphical Perception
There are two different ways in which we extract information from graphs. The first of these occurs when we take a quick glance at a graph. Without any apparent conscious effort it is possible to extract a good deal of information. Impressions such as “the graph slopes upwards” are obtained in this way. Because there is no conscious effort involved, this kind of visual processing has been called pre-attentive vision. Since our interest here is specifically with statistical graphics, we will use the term graphical perception suggested by Cleveland. The second kind of viewing is a more extended process where we consciously think about particular aspects of the graph. It is this kind of viewing that enables us to draw conclusions such as “the tallest peak in the graph occurs very close to x 4” or “a straight line through the points intersects the y axis at about y 5”. We will use the term graphical cognition for this kind of viewing. Both these ways of looking at graphs are important, but we will concentrate more on the first of them, because it is the one which makes the use of graphics attractive. In presentation graphics, an understanding of graphical perception can help us to provide agraph with what has been called inter-occular traumatic impact1. In exploratory work,such an understanding can help us develop and use techniques with the best chance of revealing hidden data features. Statistical graphs almost always encode one or more sets of numbers so that the brain’s built-in graphical perception facilities can be used to process them.If we are to have confidence in what we perceive in a plot we need to know that we can effectively decode the information in a plot. We will refer to the decoding process as a graphical perception task. Flmaigno
9
Geazlle
10
Gariffe
11
Word jumble hoax debunked I’ve previously talked about the Aoccdrnig to rscheearch at Cmabrigde uinervtisy hoax. The study described in that hoax has recently been carried out by a team at the University of Massachusetts at Amherst and the University of Durham. The data conclusively demonstrates that the hoax is incorrect. The hoax claimed that transposing letters within a word did not slow reading performance because we recognize words as whole shapes. The team led by Keith Rayner found that all kinds of letter transpositions slow reading speed. Transposing internal letters as shown in the original hoax resulted in a reading speed decline from 255 words per minute (wpm) to 227 words per minute. Performance wasworse if the transposition included the beginning or final letters of a word.
Example Sentence
Normal The boy could not solve the problem so he asked for help. Internal letters The boy cuold not slove the probelm so he aksed for help. Final letters The boy coudl not solev the problme so he askde for help. Beginning letters The boy oculd not oslve the rpoblem so he saked for help. Additionally this study examined readers’ eye movements while reading these different conditions. They found that readers needed to spend more time fixating on words in the transposition conditions and made more regressive saccades. This study only looked at letter transpositions of a single position, like the kinds used in the original hoax. I can only speculate how dramatically reading speed would be hurt with more dramatic transpositions like: The boy cluod not svloe the pelborm so he aeksd for help. Hopefully this study puts the hoax to rest. This and many other studies have made it clear that we don’t recognize words by whole shapes, but use letter information to recognize words.
Hrose
12
INFORMATION RETRIEVAL OF JUMBLED WORDS Venkata Ravinder Paruchuri Computer Science Oklahoma State University, Stillwater, OK-74078 venkarp@okstate.edu Abstract It is known that humans can easily read words where the letters have been jumbled in a certain way. This paper examines this problem by associating a distance measure with the jumbling process. Modifications to text were generated according to the Damerau-Levenshtein distance and it was checked if the users are able to read it. Graphical representations of the results are provided. Introduction Aoccdrnig to a rscheearch at Cmabrigde Uinervtisy, it deosn’t mttaer in waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht the frist and lsat ltteer be at the rghit pclae. The rset can be a toatl mses and you can sitll raed it wouthit porbelm. Tihs is bcuseae the huamn mnid deos not raed ervey lteter by istlef, but the wrod as a wlohe. The above text has circulated on the Web for several years to show how powerful the human mind is in making sense of jumbled spellings. It may be viewed from the perspectives of joint error correction and coding [1] that is done simultaneously and automatically by the mind, or from the point of view of approximate string matching [2]-[6]. It has been proposed that the human brain is able to read the words even when they are jumbled because of the following properties 1. The grammatical structure of the sentence is not disturbed in the above sentence, that is the small words [of 2 or 3 letters] or the function words [by, the, is etc] are not jumbled. Since the grammatical structure is preserved, the user is able to predict the next word in the sentence. The jumbled text not only preserves the grammatical structure, it leaves almost 45-50% of the words correct (In the above paragraph that we took 46% of the words are unchanged. 2. People generally tend to notice the first and last letters more easily than they tend to observe the middle letters. So there is less possibility of finding errors in the middle letters than the initial and last letters. 3. Although the words are jumbled in the paragraph, the jumbled words are not new words, thus making the task of the reader easier. Leporad
13
Loin
14
Mooes
15
4. The sound of the original word is preserved in the jumbled words. This also makes reading easy as people tend to read the word by its sound. 5. People read the jumbled text because of the context of the sentence. The two things that interested me, in this paper, are the use of function words and the context that plays a part in guessing the next word in the sentence. I have decided to remove the function words from the paragraph and then use the same jumbling technique to study the effect of this change. Also, to break the context of the sentence, I have taken 100 independent words that are commonly use in everyday life and then applied the jumbling technique. Approximate String Matching Approximate string matching is the technique of performing string matching to the pattern of text. The match is measured in the number of operations that are performed to match the exact string. The most common operations that are performed to match the string are insertion, deletion and substitution. The number of operations performed is measured in terms of edit distance. Examples of the operations are shown below: Insertion: monkey- monkeys Deletion: monkey- money Substitution: monkey -donkey All the above operations the number of edit distances performed are one. Some string matchers also consider transposition of two adjacent letters in the string. Transposition: lost -lots Approximate string matching has applications in many fields. Some examples are recovering the original signals after their transmission over noisy channels, finding DNA subsequences after possible mutations, and text searching where there are typing or spelling errors. Most approximate string matchers assume same cost for all the operations performed in string matching, but some matchers do assign different weights to different operations. A more detailed description about edit distance and distance functions are explained in the distance measures section. Distance measures Edit distance is the number of operations performed to transfer one string into anothOsitrch er string. There are different ways of performing the edit distance such as Levenshtein distance, Damerau-Levenshtein distance, Hamming distance, Jaro-Winkler distance, Longest common subsequence problem etc. 16
Levenshtein distance is a metric used to measure the difference between two sequences. This measure between two strings is defined by the number of edit operations used from transforming one string to another. The edit operations may be insertion, deletion and substitution of a single character. Here all the operations cost one unit. Levenshtein distance has a wide range of applications in areas such as spell checkers, dialect pronunciations and used in software’s for natural language translations [6]. As example the Levenshtein distance between Sunday and Monday is 2. Sunday ->Munday (substituting M for S) ->Monday (substituting O for U). Damerau-Levenshtein distance is similar to Levenshtein distance except that it includes an extra edit operation called the transposition of adjacent letters. Here all the operations also cost one. Damerau-Levenshtein distance has its applications in fields of fraud vendor name detections, where it can detect the letter that has been deleted or substituted, in DNA, where the variation between the two strands of DNA can be found out by this distance [6]. Hamming distance allows only substitution of letters, which cost one unit. It is applied only to the strings of similar length. It is applied in error detection and correction [6]. In this paper we apply the Damerau-Levenshtein distance to the words to find its ef1 fect on reading. This is because Damerau-Levenshtein distance has all the possible edit operations that can be performed.
Owl
17
Ram
18
Rhnicoreos
19
Experiment and Analysis: In all the experiments that I have conducted, I recorded the time each of 10 readers took to read the text. This time was then averaged. I. Removal of function words In this section I considered the actual paragraph and then removed all the function words from the paragraph to find the effect on the reader. Actual sentence According to research at an English university, it doesn’t matter in what order the letters in a word are, the only important thing is that the first and last letter is at the right place. The rest can be a total mess and you can still read it without problem. This is because we do not read every letter by itself but the word as a whole. After jumbling Aoccdrnig to a rscheearch at Cmabrigde Uinervtisy, it deosn’t mttaer in waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht the frist and lsat ltteer be at the rghit pclae. The rset can be a toatl mses and you can sitll raed it wouthit porbelm. Tihs is bcuseae the huamn mnid deos not raed ervey lteter by istlef, but the wrod as a wlohe. Without function words According research English university doesn’t matter what order letters word only important thing first last letter right place. Rest total mess still read without problem. This because read every letter itself word whole. After jumbling the above paragraph without function words Accdroing resaecrh elgnsih uvinsreity deosn’t mtaetr what order letrets wrod olny iopmrtant tnihg frist lsat lteetr rihgt plcae. rset tatol mses sitll raed whtiuot pborelm. Tihs baceuse raed every lteter istlef word wohle. Shrak
20
Does the huamn mnid raed wrods as a wlohe? Jonathan Grainger1 and Carol Whitney2 1 2
LPC-CNRS, University of Provence, 13621 Aix-en-Provence, France Department of Computer Science, University of Maryland, College Park, MD 20742, USA
A recent email message about a purported experiment run phenomena. For example, the well-known Interactive at Cambridge University provides a useful illustration of Activation [5] and Dual-Route Cascaded [6] models use a some fundamental mechanisms involved in reading. The position-specific slot encoding, which is inconsistent with message demonstrates that a text composed of words these experimental results. In response to the obvious whose inner letters have been re-arranged can be raed shortcoming of these approaches, Whitney [7,8] proposed wtih qutie anazimg esae! Although some of the readability an encoding based on ordered letter pairs (e.g. the input of this email message is probably due to top-down factors ‘take’ is represented by activation of units representing made possible by the fact that almost 50% of the words are TA, TK, TE, AK, AE, and KE). These units do not contain not mixed up, we suggest that a significant part of this precise information about letter position, or about which ‘jumbled word effect’ is due to the special way in which the letter is next to which. The same units (dubbed ‘open human brain encodes the positions of letters in printed bigrams’) were later endorsed by Grainger and colleagues words. Recent research using the masked-priming tech[4,9]. Open bigrams provide a convenient computational nique has helped to elucidate the mechanisms involved mechanism for representing the relative position of letters in letter-position coding. Masked primes are briefly in a string. They are directly motivated by the relativepresented, pattern-masked letter strings, whose effects position priming results [1,2], but also provide a quite on target processing are thought to reflect fast, automatic natural explanation for effects of transposed letters [3,4], processing [1,2]. We will briefly describe two phenomena, and other key data (discussed by Whitney in [7,8]). relative-position priming and transposition priming, that According to this open-bigram approach to letterhave been observed with this paradigm and that are position coding, perception of printed words is relatively particularly relevant for understanding letter-position insensitive to letter transpositions because enough correct coding. relative-position information is present in transposed Masked-priming studies [1,2] have shown that targetstimuli. Thus, in the transposition priming experiments word recognition is facilitated when primes are composed mentioned above, the prime ‘gadren’ has 92% (11/12) of a subset of the target word’s letters (even when the overlap with the target ‘garden’, compared with 25% proportion of shared letters is quite low, and absolute, overlap (3/12) for the orthographic control prime ‘galten’ length-dependent, letter position is violated), as long as (note that in these calculations open bigrams are limited to the shared letters are in the same order in prime and a maximum of two intervening letters). According to target stimuli. That is, priming occurs only when relative classic accounts of letter-position coding, these two prime positions are respected. For example, a six-letter word conditions do not differ in terms of their orthographic such as ‘garden’ is identified more rapidly when preceded overlap with targets. by the masked prime ‘grdn’ compared to the unrelated The superiority of this approach for explaining the condition ‘pmts’, and partly changing the order of letters ‘jumbled word effect’ can be illustrated with an example (gdrn, nrdg) destroys the priming effect. This result would taken from the original email message. The jumbled appear to be at odds with the apparent ease with which we version of the word ‘only’ (‘olny’) has 50% orthographic can read the jumbled-word email message. overlap according to the coding schemes of the Interactive However, when primes share all their letters with Activation and Dual-Route Cascaded models (and the target words, priming is robust with small changes in same overlap with ‘ogry’, for example). Traditional bigram letter order, a phenomenon referred to as transposition or trigram coding schemes do even worse with 20% overlap priming. Thus, primes formed by transposing two adjacent for bigrams (when spaces are included), and zero overlap letters in a target word (e.g. gadren – garden) facilitate for trigrams. On the other hand, the open-bigram scheme word recognition compared with appropriate control calculates overlap at 83% for this specific example, and is primes [3,4]. generally quite high for the complete set of words in the Most importantly, these two phenomena have propelled email message. Values range as a function of word length a new approach to letter-position coding that we believe and the degree of transformation, for example with 80% Sheep provides a coherent account of the ‘jumbled word effect’. overlap for ‘porbelm’ (problem), 67% for ‘wlohe’ (whole), We note in passing that current computational models of and a 53% minimum for ‘bcuseae’ (because). This relatively visual word recognition cannot capture these two key high level of orthographic overlap between the transformed word and the original word guarantees a minimum Corresponding authors: Jonathan Grainger (grainger@up.univ-mrs.fr), Carol Whitney (cwhitney@cs.umd.edu). amount of correct bottom-up input for word identification. www.sciencedirect.com
21
Tuacn
22
By Vinny Mankoo