Conf. on Embedded Systems and Applications | ESA'07 |
127
Optimized Layout for Keypad Entry System Arpit Mittal Department of Computer Science and Engineering National Institute of Technology, Warangal Warangal, Andhra Pradesh., India
Abstract - In this paper, we present a critical review of existing layouts of alphabets on a mobile keypad and attempt to improvise it to facilitate user interaction for existing schemes including less-tap, predictive text entry (on T9), prefix-based disambiguation (on Letterwise) and Simkeys. The standard constrained layout is not suited for word based disambiguation (alternatively, predictive text entry) scheme since several matches exist for the same numeric combination, some of which are frequently used words. The same argument holds for Letterwise and Simkeys. The model also addresses the problem of larger keystrokes per character (KSPC) in multi-tap and movement components based on Fitt’s law. The proposed system introduces changes in the layout and dictionary software to address the above problems by use of less-tap, and grouping commonly used key combinations together for an ergonomically better design and overall provides better disambiguation accuracy for predictive text-entry. The system was designed using corpora for written English and TextSpeak (SMS language) and evaluated experimentally. The model is expected to cause a significant rise in the text input speeds of mobile phones and other embedded devices with limited text entry capabilities, leading to improved performance and better human-computer interaction. Keywords: text entry, keypad entry, mobile phones, ergonomics
1
Introduction
The telephone keypad, originally designed for the input of digits, is also being used to enter text, since the advent and popularity of short messaging service (SMS). There are approximately one billion text messages sent per day across the globe [1]. Since the size of wireless devices is decreasing each day, telephone keypads have become the medium of choice in sending messages. Several letters are assigned to each key, generally three letters per key in alphabetical order. While this layout is currently used, it is certainly not the most optimal solution. Our model seeks to improvise this layout by rearranging characters on the keypad, to obtain a near-optimal solution to text-entry that is compatible with most existing methods of text entry proposed till date.
Arijit Sengupta Department of Computer Science and Engineering National Institute of Technology, Warangal Warangal, Andhra Pradesh., India
2
Background
On a standard telephone keypad, a letter can be entered by pressing the key to which the letter is assigned, and then choosing which of the several letters is meant by some method. The most common method is known as multi-tap, where the intended letter is obtained by pressing the key multiple times, depending on which letter is intended. Multi-tap had several shortcomings including the fact that frequently used characters are not assigned the shortest keystroke sequence. Multi-tap subsequently evolved into Less-tap [2] in which lesser keystrokes were reserved for the most commonly occurring characters in the input language. Simkeys [3] attempts to extend this concept by utilizing ‘*’ and ‘#’ keys analogous to the Shift key. In the last few years, dictionary-based methods (DBMs) have been introduced to the market, and are now widespread. These methods work by matching an entered sequence of keystrokes to words in a dictionary. Their purpose is to increase text entry speed by reducing the number of keystrokes required to enter each letter. Variants of this method include T9 by Tegic Communications, iTAP™ by Motorola, Inc., and eZiText™ by Zi Corporation. Since T9 is the most commonly used DBM, evaluation has been done using the same. Yet another approach is prefix-based disambiguation, incorporated in systems such as Letterwise [4]. The above methods use the same core components, and the standard alphabet layout. These systems would remain compatible with a change in layout without any actual change in hardware. Other systems include The Numpad Typer (TNT) [5] which has a two-keystroke input sequence for each character, and TiltText [6], which uses orientation of the phone to disambiguate choices. Since the latter two methods require a different hardware and layout, we cease their discussion beyond this point.
3
Metrics
The following three hypotheses were assumed. 1.
2.
Perfect Typist Hypothesis: There were no typing, spelling, or other errors requiring time and effort to correct. No ambiguity hypothesis: All words entered were unambiguous.
128
Conf. on Embedded Systems and Applications | ESA'07 |
Perfect Dictionary Hypothesis: All words typed were included in the dictionary.
where pi is probability of each word colliding with K, if M > 1, and zero otherwise.
Under these conditions, the following metrics have been used to theoretically evaluate the best layout.
OM of a layout is defined as the aggregate of penalty (K) for each K in the layout. OM is expected to accurately model dictionary-based techniques. Our research shows that a low value of OM minimizes KSPC and maximizes DA, thus leading to an overall increase in typing speed while using dictionary-based methods.
3.
3.1
Typing speed
Typing speed is the overall indicator in deciding usability of a layout. However, the metric cannot be accurately measured, and varies from person to person. Hence we shall have to fall back to less representative metrics to characterize text entry behavior.
3.2
3.5
Fitt’s law
Fitt’s law [8] attempts to model serial fast, aimed movements. Fitt’s law defines index of difficulty (ID) as
Keystrokes Per Character (KSPC)
KSPC is a useful metric for characterizing overall text entry behavior. KSPC is the number of keystrokes, on average, required to produce each character using a given input method.
ID log 2 (A / W 1)
(2)
where A is the amplitude of movement and W, the width of the target. It is clear that lower values of A will reduce the index of difficulty, leading to better typing speeds. This will be used in our model to rearrange keys in the layout for minimal mean amplitude of movement. It may be noted that rearrangement of letters within a key was done by a combination of KSPC, DA and OM.
The ultimate goal is to attain KSPC = 1, the standard for a computer keyboard. Except TiltText, all the discussed methods have KSPC > 1. We shall attempt to show how changing the layout and retaining the original algorithm results in improving the KSPC metric to a significant extent.
3.6
3.3
Using Fitt’s law and KSPC, we can now attempt to approximate typing speed by the following formula
Disambiguation Accuracy (DA)
DA indicates the fraction of times in which the word with the highest frequency of occurrence is the one intended by the user. An optimal design is one that maximizes the DA for a given list of words. This method only applies to dictionary based methods.
3.4
Optimization Metric (OM)
We propose a new metric, OM, which tries to form a balance between KSPC and DA, while ensuring that for a dictionary containing words that occur in 85% of all actual text, we try to ensure as few collisions as possible. For a DBM with a dictionary containing 85% of words occurring in actual text, it has been observed that multi-tap (or its variants) perform as well as the dictionary-based methods, in terms of typing speed. In practice, English databases contain up to 95% of words occurring in actual text. [7] A collision is defined as a key combination for which there is more than one possible word in the dictionary. For example, the key combination 2-2-5-3 has the words able, cake, bald and calf. We attempt to penalize only the key combinations that have a collision. For a key combination K of size N and M collisions, penalty is given by
penalty(K) N u ¦ p i
(1)
Maximum theoretical typing speed
MT
a b u ID
(3)
where MT is the mean time taken for a keystroke, ‘a’ and ‘b’ are parameters which should be chosen to reflect actual performance. For a Nokia 5100 model with thumb entry, the values of ‘a’ and ‘b’ are 176 ms and 64 ms/bit respectively. [7] The mean time taken to type a character is given by MT u KSPC. The inverse of this value gives the theoretical maximum for typing speed. In practice, the typing speed is much lesser than this theoretical maximum.
4
Design
The keypad layout design had the following steps as summarized below.
4.1
Optimization for dictionary users
The popularity of dictionaries-based word disambiguation schemes is increasing day by day. A survey by Eatoni Ergonomics [9] among Finns found a 47% penetration of DBMs. Since it is a well known fact that Finns lead the world in mobile phone usage, the current Finnish trend is expected to be replicated in the world of tomorrow. Dictionaries are, therefore to be assigned the greatest importance in text-input schemes.
Conf. on Embedded Systems and Applications | ESA'07 |
The predictive keypad design problem is difficult because if more than one letter can be placed on a given key, then a given key sequence might correspond with many possible words. Shifting letters between keys can reduce the possibility of this happening. The alphabetically unconstrained version of this problem allows any letters to be placed on any keys, and the constrained version requires letters to remain in alphabetical order across all keys. Previous studies have shown that users of predictive keypad methods with unconstrained letter placement can achieve high performance, in course of time. [10] Studies by MacKenzie et al suggest that this discovery phase lasts not more than a few hundred keystrokes. [4] In view of this observation, we have resolved to go ahead with an unconstrained version of the keypad. No changes may be made to the 1-key, currently used for punctuation, the 0-key, used for space and the ‘*’ and ‘#’ keys. Only the 2 - 9 keys can be reassigned. This constrains us to an 8-key design. This ensures learn-ability and minimum cost of conversion.
129
In the keypad layout currently in use, no use of the above table has been made. In practice, this implies that nearly 11% of the time (the probability of letter E), a user has to tap twice on the keypad, while an optimum solution would be assign these most common alphabets to single tap wherever possible. The solution was to reorder the alphabets on each numeric key in descending order of probability of usage. After optimization for dictionary and non-dictionary users, the final layout was attained as shown in Table II.
The unconstrained problem has ~1.6 u 1020 solutions; a restriction of 3 to 5 alphabets per key was added to bring this down to a manageable number.
Table II Layout after Step 2 NUMERIC KEY 2 3 4 5 6 7 8 9
There must be upper and lower bounds on the mean usage of any key; as far as possible, each key should have a nearly equal probability of usage. Key combinations such as “eta” and “jxz” should be avoided.
4.2
Optimization for multi-tap users
The design goal of multi-tap optimizations was to reduce the number of keystrokes used for commonly used alphabets. This is in principle similar to less-tap. In Morse code, the same principle was used to assign small sequences for commonly used alphabets in English language. Given in Table I is a list of English alphabets and their relative probabilities. Table I Alphabets and their probabilities ALPHABET E T A, I, N, O, S H R D L U C, M F
PROBABILITY (in percent) 11.278 8.459 7.519 6.015 5.827 4.135 3.759 3.195 2.820 2.350
1.880 1.598 1.504 1.128 0.752 0.470 0.376 0.188
W, Y G, P B V K, Q J X Z
ALPHABETS ebg adm nwjq rcp ofxz suk thv ily
In calculating the above probabilities of occurrence (listed in Table II), we used a corpus of written English. Later use of a corpus of SMS language (English) yielded the same results.
4.3
Further optimizations (using Fitt’s law)
The following heuristics were used in further optimizations of the keypad layout. 1.
Ease of use: Most mobile phone users use mobile phones with the thumb of the right hand. While two finger typing is not unknown, a very small section of people actually use this method, as it requires both hands free. The keys must be arranged so that commonly used key combinations must be placed near each other. This is in accordance with Fitt’s Law.
2.
Learn-ability: For a keypad to be accepted, it must be learnable to novice users. It is always hard to accept changes. We shall discuss the learn-ability with respect to three classes of users – new users, novice users and
130
Conf. on Embedded Systems and Applications | ESA'07 |
expert users. Classification is with respect to exposure to the standard layout. Testing has also been done using these three classes of users and their results are summarized subsequently. For the first heuristic, a weighted mean was found out of the number of cases of letters of one group occurring near the letters of another group. As an example, it was noted that the probability of occurrence of “thv” and “egb” as adjacent letters in English had a very high probability. These combinations needed to be grouped near each other.
For the second heuristic, it was important that vowels and commonly used alphabets retain their numeric positions. This was, in several cases, contradictory to the requirements of the first heuristic. In such cases, an optimum solution needed to be found. In the final layout, A, E, I and S retain their original positions, while nearly 60% of the alphabets are displaced by at most one position.
5
Proposed layout
The final layout is as shown in Figure 1.
Figure 1 Keypad Layouts The proposed layout has the following features. 1.
For dictionary users, the keypad layout has been altered to allow minimum ambiguity for any typed key combination.
2.
Multi-tap has been replaced by a variant of less-tap which requires fewer keystrokes for the most commonly used alphabets in the English language.
3.
The SMS corpus in use was compiled using actual text messages sent in Singapore and contain 7,325 words. The corpus was cleaned by removing punctuation and capitalization. However, a collection of messages in Singapore is not representative of overall class of English SMS. Hence, instead of relying solely on the SMS corpus, we have also considered BNC and Brown Corpus. The basic software model is given in the algorithm below.
Commonly used letters are grouped on nearby digits on the keypad, unless they occur on the same digit.
1.
Words and their probabilities are read from a dictionary and a table is created for the same.
The model is specifically optimized for English, but the model used can easily be replicated in any other language and used to generate optimized keypad layouts for other languages. It is to be noted that the model is not limited to English or Latin-based languages, but English has just been used as a test case for implementation and evaluation of the model.
2.
Initial keypad layouts were fed into the program, which analyzed the layout and created a word combination hash tree, for efficient program execution. Speed of execution was given priority over space complexity.
3.
Genetic Algorithm was used to study the hash trees so formed. The standard keypad layout was taken as the parent. Mutation involved reassigning a random character to any arbitrary key and crossover was done in the usual pattern.
4.
At the end of 1792 iterations, our genetic algorithm converged to final solution.
5.
Subsequently, the usage of each letter was normalized and letters were placed in order in each group.
6
Implementation details
The mobile keypad was simulated in software. The keypad layout was kept variable. Alphabet probabilities (in usage) were taken from the Online Oxford reference [11]. The following corpora were used: x SMS corpus by How, Y. [12] x British National Corpus (BNC) [13] x Brown Corpus [14]
Conf. on Embedded Systems and Applications | ESA'07 |
131
It is to be noted that since genetic algorithm was used in the initial analysis, it is evident that only a small section of the possible layouts were taken into consideration. Due to this constraint, the probability of a better layout derived from an exhaustive search, albeit small, is not zero.
7
Suggested Layout Multi-tap Dictionary
1.461 + 1.009 *
Table II Evaluation of DA, OM
Evaluation
Testing of the final layout was done in three ways. x x
7.1
Standard Layout Suggested Layout
Theoretical Evaluation User Testing
Theoretical evaluation
Each of the metrics listed above were calculated. The results are listed in Table IV.
KSPC
WPM
2.149 1.442 + 1.286 1.057 * 1.207
20.8 25.1 30.5 40.6 33.7
DA 95.41% 95.18%
From the above results it can be noted that the suggested out-performs the standard layout on almost all listed metrics.
7.2
User testing
16 15 14 Typing Speed (wpm)
13 12 11 10 9
New User
8
Novice (MT)
7
Novice (Dict)
6
Expert (MT)
5
Expert (Dict)
4 3 Week 1
Week 2
OM 10597 1433
The preliminary phase of testing was done on a numeric keypad with keys relabeled to match the desired layout. Output could be seen on the monitor. A program was used to analyze user input patterns, accuracy and speed. The program was developed in C#.NET 2005 for Windows Server 2003. The results were verified using software on a Windows Mobile phone, developed in C#.NET 2005, and on Nokia 6270/Symbian, developed in Java. In each case, the phone keypad was relabeled.
Table I Evaluation of KSPC, Typing Speed Existing Layout Multi-tap Less-tap Simkeys T9 Letterwise
23.9 43.2
Week 3
Week 4
Week 5
Figure 2 Mean typing speed for Multi-tap test users
Week 6
Conf. on Embedded Systems and Applications | ESA'07 |
Typing Speed (wpm)
132
22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3
New User Novice (MT) Novice (Dict) Expert (MT) Expert (Dict)
Week 1
Week 2
Week 3
Week 4
Week 5
Week 6
Figure 3 Mean typing speed for Dictionary-based test users 74 people volunteered for the experiment. All of them were either university staff or students. Subjects who had no previous exposure to mobile phones were classified as new users. The mobile-literate users were classified into multi-tap and dictionary users. All dictionary users used T9. Mobile-literate users were further classified into novice and expert categories depending on typing speed. 13 people were left out of the testing as they couldn’t be classified as either novice or expert. This was done to create specific and unambiguous classification. Test messages were taken from e-books from Project Gutenberg. Experiments were continued for 6 weeks with 3 sessions of 1 hour duration each week. Performances over the week were analyzed and mean performance is presented in Figure 2 and Figure 3. Table III Distribution of Test subjects Category New Users Novice Users using Multi-tap Novice Users using T9 Expert Users using Multi-tap Expert Users using T9
8
Count 12 22 9 6 7
refers to the layout used in the keypad design. In our case, the Discovery phase was found to take as much as 2 weeks of testing. Thus the system has questionable learn-ability. In the motor-reflex acquisition phase (which lasts for thousands of keystrokes) speed of input increases logarithmically. It can be seen that in this phase, mobileliterate users perform better than new users. The terminal phase, the Fitt’s Law phase, pertains to advanced stage of learning. While this is only asymptotic to reality, it is an important approximation made by theoretical models. Here, the keypad geometry and frequency with which pairs of keys are operated in succession determine the overall entry speed. It is here that our keypad layout is expected to be the clear winner in terms of text entry speed. Use of alphabetically constrained layout ordering, while easier to learn, effectively sets a lower upper bound on the maximum typing speed available. Despite the fact that the now-standard QWERTY layout might actually be difficult for any novice user to get used to, it significantly increases the maximum typing speed possible. A keyboard layout of the form ABCDEF would, in effect, be easier for any novice user to find keys in, but it would be overkill in the long run, slowing down typing speeds significantly.
Discussion
Recently MacKenzie et al suggested that learning is divided into three phases [4]. In the Discovery phase, which lasts for a few hundred keystrokes, the speed of entry is dominated by user’s familiarity with convention. This
The proposed layout faces similar problems of acceptability. The layout might be difficult to learn, but by constant usage, the user becomes an expert in text input, and subsequently typing speeds rise.
Conf. on Embedded Systems and Applications | ESA'07 |
9
Future Work
The proposed model is built for English and other Latin-based languages. Work is currently on to support Indian languages which have conjuncts in addition to alphabets. We have already built a model for the languages Hindi, Bengali, Gujarati and Telugu with their respective scripts and Hindi/Bengali with iTrans. [15] Work on languages such as Chinese or Japanese, where there is no concept of a word, is also on the anvil.
10 Conclusion The proposed layout has immense potential for use in text intensive applications. In Multi-tap mode its performance competes with Less-tap while it performs at least 8% better than the best T9 solution available in terms of keystrokes per character. The popularity of these keypads may well depend on the popularity of mobile phones. It is clear that mobile phones are expected to grow in numbers and features, with several non-voice features like cameras, word processors and database management systems. On the other hand, computers are decreasing in size, with laptops, notebooks, palmtops and whatever comes next. The final hybrid solution will be neither a phone nor a computer, as we know them today, but a multipurpose embedded device with a small display and smaller keypads In such a scenario comes in the mobile keypads of today, optimized to the fullest with the proposed layout scheme, to service the text input needs of these phones before voice recognition becomes efficient enough to take over.
133
[3] Ha R. W. K, Ho P. and Shen X. S., “SIMKEYS: An Efficient Approach in Text Entry for Mobile Communications”, IEEE CCNC pp. 687-689, 2004. [4] MacKenzie I.S., Kober H., Smith D., Jones T., and Skepner E., “LetterWise: Prefix-based disambiguation for mobile text input”, Proc. UIST 2001, ACM Press (2001), pp. 111-120 [5] Ingmarsson, M., Dinka, D., and Zhai, S., “TNT – A numeric keypad based text input method”, Proc. CHI, ACM Press, pp. 639-646, 2004. [6] Wigdor, D., and Balakrishnan, R. “Tilt text: Using tile for text input to mobile phones”, Proc. UIST 2003, ACM Press (2003), pp. 81-90. [7] Silfverberg, M., MacKenzie, I. S., Korhonen, P., “Predicting Text Entry Speed on Mobile Phones” Proc. ACM Conference on Human Factors in Computing Systems CHI 2000, ACM, pp. 9-16, 2000. [8] P.M. Fitts, “The information capacity of the human motor system in controlling the amplitude of movement”, Journal of Experimental Psychology Vol. 47, pp. 381-391, 1954. [9] Gutowitz H., “Barriers to Adoption of DictionaryBased Text-Entry Methods: A Field Study”, Eatoni Ergonomics Research Papers, pp. 1-8, 2003. [10] James C.L., and Reischel K.M., “Text input for mobile devices: Comparing model prediction to actual performance”, Proc. CHI, ACM Press, pp. 365-371, 2001.
11 Acknowledgements The authors would like to acknowledge the Project Gutenberg Literary Archive Foundation, Carnegie-Mellon University for the collection of books, the Brown Corpus and Oxford Online Reference for linguistic data. The authors are also grateful to the staff and students at National Institute of Technology, Warangal, who devoted their time and energy for evaluation of the results and contributed with valuable suggestions for improvement of the model.
[11] Oxford Online Reference .
homepage,
[12] How Y. and Kan M., “Optimizing predictive text entry for short message service on mobile phones”. Proc. HCII 05, 2005. [13] Leech, G., Rayson, P., and Wilson A., “Word Frequencies in Written and Spoken English: Based on the British National Corpus”, Pearson ESL (2001).
12 References
[14] Brown Corpus, Brown University (1979).
[1] Buckingham S., “An introduction to the short message service”, Mobile Lifestreams Limited, 2000.
[15] iTrans Homepage .
[2] Pavlovych A. and Stuerzlinger W., “Less-Tap: A Fast and Easy-to-Learn Text Input Technique for Phones”, Proc. Graphics Interface, 2003.
