1 Introduction

Mobile computing has become the norm rather than the exception. In May of 2003, the New York Times reported that sales figures of laptops overtook desktops for the first time[1], and as of May 2005, laptops accounted for more 53% of the total US PC market[2]. At the same time, handheld devices have continually increased in popularity since the inception of stylus-based PDAs (Personal Digital Assistants) of the early 1990s and are more popular than ever today, with over 1.15 billion mobile phones sold in 2007 alone[3]. By the end of 2008, it is projected that over two trillion text messages will have been sent[4]. These developments, together with the increased sophistication and diminishing sizes of popular devices such as digital cameras and portable media players, suggest that it has become vital for their user interfaces to not just be afterthoughts. If a user feels it takes too long to enter someone’s contact information into his phone and seeks out a paper and pen, then that device has failed. If a user plugs a mouse into her laptop rather than using the built-in touchpad, then that device is not living up to its potential. If a user waits until she gets home to her desktop PC to reply to an email rather than using the device in her pocket, then there is work to be done. This thesis attempts to improve upon current methods of interaction with mobile devices in three parts. Part one deals with touchpads as found in many laptop and notebook computers and can be extended to apply to the touchpads found in devices such as the Apple iPod (“Click Wheel”) and Microsoft Zune (“Zune Pad”). Additionally, since touchscreens in devices such as the Apple iPhone and Palm Treo are simply transparent touchpads affixed on top of displays, the methodology described in this section also applies to these devices. Part two attempts to address speed and accuracy issues in entering text onto touchscreen devices by means of a stylus. Devices such as tablet PCs or Palm Treos are examples of devices in this category. The third and final part of the thesis introduces a new way of entering text onto a small touchscreen device without the use of a stylus (i.e. using one’s finger). Specific focus is on the case of a mobile phone with an on-screen numeric keypad.

1.1 Touchpads

Touchpads have become the pointing device of choice for notebook computers. Since notebooks are often operated in constrained spaces such as the seat-back trays of economy flights or the unwieldy platform of one’s lap in the back of a taxi, a mouse cannot always be effectively used. Hence, alternatives were devised. Early popular pointing devices on laptop computers were clip-on trackballs which later led to integrated trackballs and isometric joysticks such as the TrackPoint on ThinkPad brand notebooks. Apple introduced the touchpad (the “Trackpad”) in their PowerBook 500 series notebook computer in 1994. Their device implemented selections via a physical button. A “lift-and-tap” method of clicking was added soon after and the touchpad has remained relatively unchanged until the introduction of the MacBook Air with its “Multi-Touch Gesture Pad” in 2008. However, even the Multi-Touch touchpad still offers only lift-and-tap or a button as choices for selection.

1.1.1 Technology

There are two types of touchpads in general use today: capacitive and force-sensing-resistive (FSR). Capacitive touchpads are found in devices such as the Apple iPhone (see Figure 1b), Apple MacBook, most notebook computers, and the Tactile Touchpad prototype described in this thesis. The main quality that distinguishes them from the alternative FSR technology is that an electrically conductive object (most commonly a finger) is required to use it. A plastic pen or stylus will not work, no matter how hard you press, but a metal spoon will. FSR touchpads are used in products such as the iRiver H-10 media player (see Figure 1a), the Siemens EMOTY CX70 mobile phone and the RollPad prototype described herein. A stylus, a finger, or almost any other object that can apply direct pressure to its surface, can be used with it. However, due to the necessity of having direct pressure applied to its surface, it will not function if something rigid, such as glass is placed over it, therefore precluding its use in devices such as the Apple iPod Touch which use scratch resistant glass to protect the touchscreen.

(a) (b)

Figure 1 (a) the iRiver H-10 media player with Force-Sensing-Resistive touchpad, and (b) the Apple iPhone with Capacitive touchpad

A third type of technology called electro-magnetic resonance was used for the UniPad method described in this thesis. This technology is most commonly found in pen tablet devices, for example, those made by Wacom[5]. The most distinguishing feature of the technology is that a special pen or mouse which contains a resonant circuit must be used with the tablet. A finger, or any other object, will not operate the device. As with capacitance touchpads, a glass or other rigid materials may be placed on top of it, and the device will still operate – because actuator proximity rather than direct pressure is required for operation. There are also other touchpad and tablet technologies currently available such as those found in digital whiteboards and table computers, but these are outside the scope of this thesis.

Figure 2: Wacom’s Intuos3 Pen Tablet utilizing Electro-Magnetic Resonance technology

Capacitive touchpads work on the principle of coupling capacitance. Two layers of measurement electrodes in the surface of the touchpad sense when the capacitance between them changes. A conductive object, such as a finger, causes the capacitance between the layers to change when it comes into close proximity or makes contact with the touchpad. To detect the exact location a grid is formed with one layer of electrodes arranged vertically and the other layer arranged horizontally. One set of electrodes continually send out a signal which the other set continually senses. When a finger approaches, the capacitance of the closest electrode intersections is lowered, and thus, the location can be determined. Then, for example, the direction of finger movement may be computed by comparing previous locations of lowered capacitance to the current location.

Force-sensing-resistive touchpads work by utilizing a property of a special film whereby, when pressure is applied, its electrical resistance is decreased. Two layers of this film are used: one for sensing horizontal position and the other for sensing vertical position. Together the exact location of contact can be determined. Pressure levels can also be determined by measuring the amount of resistance at the point of contact. The lower the resistance, the greater the pressure being applied.

Electro-magnetic resonance tablets work by having the tablet emit a magnetic field inducing energy into the resonant circuitry in a special mouse or stylus. In this way, no batteries are required to power the electrical circuit in the mouse or stylus. The device then in turn uses this energy to return a magnetic signal back to the tablet. Precise and detailed data can be derived from this arrangement; not only is the horizontal and vertical position available, but also, in the case of the stylus, the angle and pressure level as well.

1.2 Touchpads vs. Mice

Although touchpads are also available for desktop computers, most people prefer a mouse. So, why is a mouse a better pointing device than a touchpad when space is not an issue? The answer may lie in the separation of selection from positioning. Using a mouse, the pointer is positioned by moving the mouse on a mousepad. The device is gripped between the fingers and thumb and movement occurs via the wrist and forearm. With a touchpad, pointer movement is accomplished by sliding a finger along the touchpad’s surface. Both are generally used as “relative positioning” devices, where the on-screen pointer moves relative to its previous position when the device or finger moves.

For a mouse, selecting is the act of pressing and releasing a button while the pointer is over an icon or other screen object. Double clicking and dragging are related operations that also require pressing a button. There are two common implementations for selecting with touchpads: (a) using physical buttons, or (b) using lift-and-tap. Both inherit problems that are trying to be corrected in the Tactile Touchpad.

1.3 Physical Buttons

Most touchpads include physical buttons that are typically operated with the index finger or thumb. If an index finger is used, the finger must move frequently between the touchpad and the buttons and this impedes performance compared with the same procedure using a mouse. If the thumb is used, then positioning and selecting proceed in concert, as with a mouse; however, the result may be sub-optimal because of interference between the muscle and limb groups engaged. A similar problem has been noted for trackballs (MacKenzie, Sellen, & Buxton, 1991), wherein high error rates (particularly for dragging tasks) are attributed to the “closeness” of the muscle and limb groups required for the separate acts of positioning and selecting. With a mouse, on the other hand, positioning occurs primarily via the wrist and forearm, while selecting occurs primarily through the fingers. Thus, the limbs and muscle groups are separate for each task and tend not to interfere.

1.4 Lift-and-Tap

Because of the problem noted above, most touchpads also support “lift-and-tap” as an alternative to pressing buttons. However, this is perhaps replacing one problem with another. This is illustrated by considering the basic transactions with computer pointing devices. According to Buxton’s three-state model of graphical input (Buxton, 1990), these can be modeled by three states:

State 0 out-of-range (the device/finger is elevated)
State 1 tracking (pointer movement)
State 2 dragging (movement with button depressed)

These are identified in Figure 3 annotated for mouse interaction.

Figure 3: Buxton’s three state model of graphical input
with labels appropriate for mouse interaction (Buxton, 1990)

For touchpads and mice, pointer motion occurs in state 1, the tracking state. The comparison becomes interesting when one considers the state transitions required for clicking, double clicking, dragging, and clutching. (Clutching is the act of lifting the mouse or finger from the mousepad or touch surface and repositioning it.) Table 1 identifies the state transitions for the most common operations for a mouse and a lift-and-tap touchpad. A few observations follow. In general, operations require more state transitions with a lift-and-tap touchpad than with a mouse. A simple click on a mouse begins and ends in state 1, whereas on a touchpad it begins in state 1 and ends in state 0. To return to pointer positioning (state 1), the finger must resume contact with the pad, and if this occurs too quickly a dragging operation occurs. Note as well that clutching on a lift-and-tap touchpad is confounded with clicking and dragging. This is not the case with a mouse.

Table 1: State Transitions for Common Operations
Using a Mouse and a Lift-and-Tap Touchpad.

Operation	Mouse	Lift-and-tap Touchpad
Pointer Positioning	1	1
Single Click	1-2-1	1-0-1-0
Double Click	1-2-1-2-1	1-0-1-0-1-0
Dragging	1-2	1-0-1-0-1
Clutching	1-0-1	1-0-1

Figure 4 compares dragging for a mouse and touchpad. Two observations follow: (1) lift-and-tap necessitates extra state transitions when compared to a mouse, and (2) the use of state 1-0-1 transitions is confounded with clutching (not shown) which uses the same state transitions.

a) carroll-f2a b) carroll-f2b

Figure 4: State transitions for dragging. (a) mouse (b) touchpad using lift-and-tap.
The first horizontal bar is the cursor positioning before dragging.

In the following sections, each of the three prototype devices created for this thesis are described. Afterward, an evaluation of each is presented.

2 The Tactile Touchpad

In view of the preceding discussion, it is worth exploring alternate, perhaps better, implementations for state transitions. One possibility is to implement them by pressing harder with the pointing/positioning finger. A mouse button provides aural and tactile feedback when it is pressed, and this is an important component of the interaction. Similar feedback may be elicited from a touchpad by means of a mechanical solenoid or relay positioned under the pad and activated with an electrical signal to create a “click” sensation in the fingertip. Since a mouse button clicks both when pressed and when released, the same response is desirable for a Tactile Touchpad to achieve a more natural feel.

To prevent spurious clicks, the transitions should include hysteresis. That is, the pressure level that maps to the button-down action of the transition from State 1 to 2 should be higher than the pressure level that maps to the button-up action of the transition from State 2 to 1. This is illustrated in Figure 5. The correct thresholds must be determined in user tests.

Figure 5: Pressure-state function. A click is generated for
state 1-2 transitions and for state 2-1 transitions.

There is prior work on embedding a solenoid under a mouse button to create tactile feedback. A study by Akamatsu and MacKenzie (1996) found significant reductions in movement times for target selection tasks using a modified mouse incorporating tactile feedback as compared to an unmodified mouse. Using a Fitts’ law analysis of the data, it was found that the tactile condition produced the highest throughput of all tested conditions. It was surmised that similar results would be achievable with the Tactile Touchpad. One can provide aural feedback through the computer’s existing sound system. However, it is felt that the combination of spatially-placed aural and tactile feedback at the finger tip is preferable to spatially-displaced audio-only feedback using the system’s loudspeaker, although the latter is worthy of investigation.

The Tactile Touchpad is illustrated in Figure 6. For the prototype, a hole was cut in the bottom of a Synaptics T1002D capacitive touchpad and a Potter & Brumfield T90N1D12-5 relay was installed. A wooden platform attached to the base provides space for the relay. The relay is controlled by signals sent from the host’s parallel port.

The Synaptics touchpad includes an x-y-z mode in which the z-axis information is the applied pressure. The software uses z-axis information to determine when to energize and de-energize the relay.

Via informal tests with pilot participants, it was determined that, of the 256 pressure levels detected by the touchpad, a value of 140 with a hysteresis value of 5 produced the best response - one similar to the feel of a physical mouse button.

(a)

(b)

Figure 6: The Tactile Touchpad. (a) top view. (b) underside view.

2.1 ISO Testing of Pointing Devices

Although there is an abundance of published evaluations of pointing devices in the disciplines of human-computer interaction and human factors, the methodologies tend to be ad hoc, and this greatly diminishes the ability to interpret the results or to undertake between-study comparisons. Fortunately, there is an ISO standard that addresses this particular problem (ISO, 1999). The full standard is ISO 9241, “Ergonomic design for office work with visual display terminals (VDTs)”. The standard is in seventeen parts. Part 9 of the standard is called “Requirements for non-keyboard input devices”.

ISO 9241-9 describes, among other things, quantitative tests to evaluate computer pointing devices. The procedures are well described and allow for consistent and valid performance evaluations and comparisons of pointing devices.

The standard quantitative test is a point-select task. The user manipulates the on-screen pointer using the pointing device and moves it from a starting position to a target and selects the target by pressing and releasing a button on the device. There are many variations on this test; however, a simple reciprocal selection task is easiest to implement and allows for a large quantity of empirical data to be gathered quickly. The task is “reciprocal” because the user moves the pointer back and forth between targets, alternately selecting the targets. The selections are “blocked” with multiple selections per task condition.

As the point-select task is carried out, the test software gathers low-level data on the speed and accuracy of the user’s actions. The following three dependent measures form the basis of the subsequent quantitative evaluation:

Movement Time. Movement time (MT), or task completion time, is the mean time in seconds or milliseconds for each trial in a block of trials. Since the end of one trial is the beginning of the next, the movement time is simply the total time for a block of trials divided by the number of trials in the block.

Error Rate. Error rate (ER) is the percentage of targets selected while the pointer is outside the target.

Throughput. Throughput (TP) is a composite measure, in bits per second, based on both the speed and accuracy of performance. The measure was introduced in 1954 by Fitts (1954), and it has been widely used in human factors and experimental psychology ever since. See (MacKenzie, 1992; Welford, 1968) for extensive reviews.

Throughput, as specified in the ISO standard, is calculated as follows:

Equation 1: The equation for throughput

where

Equation 2: The equation for effective index of difficulty

The term ID_e is the effective index of difficulty, and is measured in bits. It is calculated from D, the distance to the target, and W_e , the effective width of the target.

The term MT is the movement time to complete the task, and carries the unit seconds. Thus, throughput carries the unit bits per second, or just bps.

The use of the effective width (W_e) is important. W_e is the width of the distribution of selection coordinates computed over a block of trials. Specifically,

Equation 3: The equation for effective width

where SD_x is the standard deviation in the selection coordinates measured along the axis of approach to the target. This implies that W_e reflects the spatial variability or accuracy that occurred in the block of trials. As a result, throughput is a measure that balances the tradeoff in the speed and the accuracy of the user’s performance. There is recent empirical evidence that the reciprocity between speed and accuracy in the computation of throughput is, in fact, equal (MacKenzie & Isokoski, 2008). In some sense, throughput reflects the overall efficiency with which the user was able to accomplish the task given the constraints of the device or other aspects of the interface.

It is important to test the device on difficult tasks as well as on easy tasks; so, multiple blocks of trials are used, each with a different target distance and/or target size.

In the following section, another application of touchpads is presented, along with a new design selected for empirical evaluation in this thesis.

3 UniPad

The push toward mobility in computing has rekindled interest in one of the oldest areas in office automation: text entry. This is due in no small measure to the huge popularity of SMS messaging on mobile phones. Related services on other mobile devices are also gaining popularity. In view of this, researchers and companies are ambitiously searching for improvements to mobile text entry techniques; improvement that are easy to learn, easy to use, and capable of high speed entry of machine-readable alphanumeric text.

Viable entry methods depend on context. Two broad categories are physical key-based and finger and stylus-based devices. The former include most mobile phones and pagers, the latter include personal digital assistants (PDAs), tablet PCs, portable media players and large screen phones. Stylus and finger-based input is further divided between handwriting with automatic recognition and tapping on soft, or virtual, keyboards. The focus of this section of the thesis is stylus-based text entry using handwriting recognition.

This paper presents the design and evaluation of UniPad. UniPad combines single-stroke handwriting recognition with language-based acceleration. Features promoting high-speed entry are movement minimization and reduced attention demand. These are explained in the discussions that follow, after which an evaluation is presented.

3.1 Entry Speed

Mobile text entry speeds are unlikely to rival those for desktop touch typing, where rates of 50 to 70 wpm are common[6]. However, such rates are rarely achieved without years of practice.

Numerous text entry techniques exist both commercially and as research prototypes. Although provocative, comparing entry rates among techniques is misleading unless framed in the proper context, such as the type of task, or users’ prior computing experience, or practice with the technique. Soft keyboard entry rates, for example, have been measured at 40+ wpm (MacKenzie & Zhang, 1999) and even 70+ wpm[7], but these rates were achieved following considerable practice. When users first confront a soft keyboard, entry rates are much lower, on the order of 18–28 wpm for a Qwerty layout (Fleetwood et al., 2002; MacKenzie & Zhang, 1999; MacKenzie, Zhang, & Soukoreff, 1999) or as low as 5–7 wpm for an unfamiliar layout or method (MacKenzie & Zhang, 2001; MacKenzie et al., 1999; Oniszczak & MacKenzie, 2004).

Handwriting speeds are in the 13–22 wpm range (Card, Moran, & Newell, 1983). However, when coupled with recognition, the interaction is quite different, since users must attend to and respond to the on-going process of recognition and error correction. As well, most implementations impose constraints on users’ writing style to match the requirements of recognition (e.g., stroke sequence and/or direction, or stylized strokes). Achieving higher speeds with handwriting recognition, therefore, necessitates new strategies, such as using a simplified stroke alphabet or reducing the number of required strokes using word completion. UniPad includes both these features, as now explained.

3.2 Simplified Strokes

An 1895 report on typewriting indicated that “the strain of sight typing on the eyes and mind increases with speed, until a point is reached where it cannot be kept up for any length of time” (quoted from (Greenberg, Darragh, Maulsby, & Whitten, 1995)). And so, the skill of “sight typing” evolved into “touch typing” where typists operate the apparatus without attending to the keys. However, a similar migration of skill is generally not possible with soft keyboards, since the user must maintain visual fixation on the keys. Consequently, the attention demand is substantial.

Is eyes-free handwriting possible? Generally no, due to cumulative errors in registration and stroke alignment. However, an interesting variation does support eyes-free entry via single-stroke-per-symbol text input by use of a stylus (Goldberg & Richardson, 1993). Examples include Unistrokes and commercial implementations such as Graffiti and Jot (also known as Graffiti 2), which are found on many PDAs. Typically, the single strokes are stylized to mimic the symbol’s character-equivalent shape while conforming to the one-stroke-per-symbol constraint. Figure 7 shows the Graffiti stroke alphabet used with UniPad. Most strokes closely match either the uppercase or lowercase Roman letter. Although differences remain, the alphabet is quickly learned – with about 97% accuracy after five minutes of practice (MacKenzie & Zhang, 1997).

Figure 7: Single stroke alphabet example

Unlike soft keyboards, single-stroke text entry afford eyes-free entry. Strokes can be made anywhere on the digitizing surface. Successive strokes are spatially independent and can occur, for example, directly on top of preceding strokes. Thus, the user may attend to other aspects of the interaction, such as the edit buffer, digital ink trails (if any), or word completion aides accompanying input.

3.2.1 Word Completion

With word completion, users attend to an ever-changing candidate list, which bears a cognitive load (Greenberg et al., 1995; Mankoff, Hudson, & Abowd, 2000). So, the benefit of eyes-free entry is mitigated by the attention required by the word completion list. A central motivation of the present research is to find a “sweet spot” – a way to combine single stroke text entry with word completion to increase throughput without overly taxing the user. Methods such as TextPlus by SmartCell Technology[8] go some way towards this goal by adding word and phrase completion to Graffiti. However, a key feature in the UniPad method is movement minimization. The single-stroke-per-symbol feature inherently reduces required stylus movement and in this method, the input and output regions are superimposed. That is, single stroke characters are entered on top of the language-based acceleration aides, thus stylus movement is further minimized. It is conjectured that attention is lessened by superimposing the digital ink on the acceleration aides.

3.2.2 Acceleration Techniques

UniPad implements several acceleration aides, described as follows.

Word completion using a candidate list

The default dictionary, based on the British National Corpus[9], has 64,566 unique words with frequencies from a total of 90,563,964 words. The average word size is 8.45 characters based on a simple mean, or 4.59 characters if weighted by word frequency.

As entry proceeds, a list of candidate words is produced on each pen-up. If the desired word appears, the user simply taps on it. This terminates entry of the current word and delivers the result to the application with a terminating SPACE.

The user may ignore the list, choosing instead to continue entering single stroke characters, or attend to the list if there is a sense that the desired word is present. Two characteristics of the candidate list can be adjusted: the size of the list, and the number of strokes before word completion begins. This is consistent with commercial systems such as WordComplete by CIC[10]. The tests use a list size of five, with word completion beginning on the first stroke. For example, the word “recognize”, normally requiring nine strokes (including a terminating space), requires just four:

( = tap word in candidate list)

The desired word appears in the candidate list after the third stroke. Tapping on it, selects the word and delivers it to the edit buffer with a terminating space.

Size Precedent Sort Order

Although the most obvious way to present the candidate list is sorted by word probability, a slightly different approach is taken here. The words are sorted first by size, and then lexically sorted within size. Thus, more information is inherently encoded in the ordering. The expected effect – although difficult to quantify – is to reduce the visual scan time to find the desired word.

Suffix Completion

Reduced keystroke suffix completion in languages such as English have a long history in assistive technologies, where common suffixes are appended typically with a single keystroke (Higginbotham, 1992). With this system, suffix completion mode is entered when a candidate word is selected with a top-left to bottom-right stroke ((). The candidate word list is replaced with a set of popular suffixes. The size of the list is programmable, however these tests use a list size of twelve: -s, -ed, -er, -est, -ly, -able, -ful, -ing, -ion, -ive, -ment, -ness. Tapping a suffix appends it and adds a terminating space.

As an example, “parts”, “parted”, “parting”, or “partly” can be entered with just four strokes (Note: “part” appears in the candidate list after two strokes):

( = tap entry in suffix list)

Suffix completion reduces keystrokes two ways. The first was just demonstrated, but, additionally, words containing suffixes are eliminated from the dictionary (with counts added to the base word count). The ancillary benefit is that the candidate list contains a richer set of possibilities. This also reduces keystrokes since a given word stem expands to a more diverse set of choices.

Although suffix completion can save keystrokes, it would seem that the primary benefit is in simplifying the interaction primitives. Straight-line strokes are fast (< 200 ms) (Goldberg & Richardson, 1993; Isokoski, 2001) and users quickly assimilate the “” pattern using chunking or unitization (Welford, 1968).

Frequent Word Prompting

At the beginning of each word the input pad is blank, as there is no word stem upon which to generate candidates. This void is filled with a list of the most frequent words in the dictionary. This is the “frequent word prompt”. The size of the list is programmable, however a size of twelve was used here. This coupled with the default dictionary yields “for”, “the”, “you”, “and”, “was”, “that”, “of”, “a”, “to”, “is”, “in”, “it”. A tap selects the word, delivering it to the edit buffer with a terminating space.

Other Features

UniPad includes several other features supporting general purpose text entry and editing. Uppercase and caps lock modes are implemented using a bottom to top (#) stroke, in a manner that mimics commercial systems, such as Graffiti on Palm PDAs. A right to left stroke (!) serves as a backspace.

Several soft buttons are included, for example to clear the edit buffer (Clear) or enter a carriage return (Enter). There is a button to bring up a soft keyboard (SoftKey) for inputting punctuation and other symbols. Dedicated character-delete (Char<=) and word-delete (Word<=) buttons are implemented to facilitate users’ editing strategies.

UniPad also includes a variety of space management techniques, to coordinate spacing for punctuation between and within words, phrases, and sentences. For example, the default space terminating a word is removed if the following character is a period.

3.2.3 UniPad Interface

The general operation of UniPad and its language-based acceleration techniques are now illustrated.

Figure 8 is a screen snap of UniPad working in concert with the text entry evaluation software. Phrases are presented to the user for input. During entry, performance measures are gathered and saved for follow up analyses. In the example, the user is entering “hours”. After two strokes, the desired word has yet to appear in the candidate list, so the user proceeds to enter the third stroke. The digital ink trails accompanying input appear directly on top of the candidate list.

Figure 8: Interaction using candidate list

In Figure 9, the user is about to enter “of”. Since this word appears in the frequent word prompt, a simple tap suffices.

Figure 10 shows interaction using suffix completion. To enter the word “parking” the user has a few options. The plain single stroke method requires eight strokes:

( = tap and pen Up)

Using suffix completion however, only five strokes are required:

( = tap entry in suffix list)

The fourth stroke is a top-left to bottom-right gesture beginning on top of the entry “park”, which appears in the candidate list with the stem “par”. This brings up the suffix completion list. A final tap on -ing completes the word and adds a terminating space.

Figure 9: Interaction using frequent word prompt

ExperimentScreenSnap-3b

Figure 10: Interaction using suffix completion

3.2.4 Keystrokes Per Character

Although isolated examples such as those above serve to explain interaction features, they fail to characterize or quantify the overall benefits in terms of interaction primitives. For this, Keystrokes Per Character (KSPC) is used (MacKenzie, 2002). KSPC is the number of keystrokes (in this case, “stylus strokes”) required, on average, to generate each character of text using a given interaction technique in a given language. For plain single stroke text entry, KSPC = 1, since each stroke generates a character. For techniques such as Multitap on mobile phones, KSPC > 1, since more than one key press is required to produce each character. However, KSPC can be less than 1 if language-based acceleration techniques are used, as just discussed.

Ideally, the lowest possible KSPC in a word completion text entry system occurs where each keystroke generates a word. With the default dictionary, the mean word size is 4.59 characters. Adding 1 for a terminating space character, the lowest possible value for KSPC is 1 / 5.59 = 0.179. (Prediction can also work at the phrase level (Darragh & Witten, 1993; Masui, 1998), but this is not explored here.)

KSPC is computed using the default dictionary with numerous settings for UniPad’s acceleration features. Figure 11 shows results for various list sizes for the candidate list (“CL”) and frequent word prompt. The calculations do not include the keystroke reduction due to suffix completion.

The point at the top-left in Figure 11 shows the result if language-based acceleration is not used: KSPC = 1. The effect of adding language-based acceleration is seen moving across and down the figure. The point at the bottom-right has KSPC = 0.428. Adding suffix completion reduces this by a very small amount; however, as noted above, the primary benefit with suffix completion is in simplifying the stroke primitives.

Figure 11: KSPC by list sizes for Frequent Word Prompt and Candidate List (CL)

The KSPC figure just cited represents a best-case scenario. That is, UniPad’s acceleration features afford English text entry with just under 0.5 keystrokes per character. However, several important issues remain. First, it is not clear that users will actually use the acceleration features effectively. Missed opportunities tend to push KSPC up, to the point where users are interacting essentially in plain single-stroke text entry mode. Clearly, learning is important as well. With practice, a user’s KSPC may fall as features are learned and exploited at the earliest opportunity. And finally, it is not clear that reductions in KSPC yield a corresponding increase in throughput, since the acceleration features also add cognitive demands to the interaction. To test for the potential benefits of UniPad, a user study is required. This user study is presented later in this thesis.

4 Rollpad

Mobile devices are generally smaller than their non-mobile counterparts and often suffer from compromises due to their size. Most mobile phones use a 12-key physical keypad. Although fine for entering digits, methods requiring extra effort are needed to enter alphabetic characters. Full QWERTY keypads have gained some popularity in phones such as the Blackberry and iPhone lines. (See Figure 12)

Figure 12: iPhone with soft keyboard on the left and Blackberry with physical keyboard on the right

However, there is a tradeoff between how small the keys can be and remain accurate to use versus how large a keypad can be and still remain pocketable and desirable to own. The iPhone attempts to address this by correcting mistyped words via a built-in dictionary. Although helpful, the system is not perfect and also introduces a requirement for the user to “trust” the device as is found in the T9 method described below. This requirement of trust takes away the ability to view one’s entered text in real time and is ultimately less satisfying an experience as it takes control away from the user. Based on personal experience with the Blackberry 6710 and the 8GB iPod Touch (which utilizes the same interface as the iPhone) it was observed that there was a greater feeling of control and accuracy while using the Blackberry keyboard over the iPod’s. However, the author does not enjoy putting the huge Blackberry into his pants pocket as compared to the svelte and narrow iPod Touch. With the ever increasing demand for mobile text input on ever smaller devices, improved alternatives are desirable.

4.1 12-Key Text Entry Methods

There are a few variations on the 12-key keypad, but most are similar to that in Figure 13. The advantage of the 12-key system is its familiarity and ability to have reasonably sized buttons even in very small devices.

TouchpadKeypad

Figure 13: 12-key Keypad Layout (with Mode key)

Early keypads omitted the letters q and z, resulting in 3 characters per alpha key. Most current keypads add q to the 7 key and z to the 9 key (see Figure 13). Occasionally, q and z are on the 1 key as was common on Australian phones[11]. This arrangement is used for the device described herein. With at least 3 characters on each alpha key, a disambiguation method is needed. Some current methods are reviewed.

4.1.1 Multitap

With Multitap, a desired character is obtained by tapping its key one or more times until it cycles into view. For example, to obtain the character c, the 2 key is pressed quickly 3 times, so the device cycles through a, b, and then c. When the desired character is obtained, a time-out or pressing a “time-out key” is required before entering another character on the same key. Problems arise when the user pauses longer than the time-out period while cycling, or does not wait long enough before entering the next character on the same key. Although widely used, the method remains slow and error-prone (Pavlovych & Stuerzlinger, 2004).

4.1.2 Dictionary Based Disambiguation

T9 by Tegic Communications uses dictionary-based disambiguation. Similar methods include iTap by Motorola and eZitext by Zi Corp. Users press keys once per character and a dictionary works to match words with key sequences. This works fine when the desired word uniquely appears in the dictionary. Where two words map to the same key sequence (e.g., cat and bat both map to 228), the device lists the entries ordered by probability. The desired word is selected using a special “next key”. If the desired word is not in the dictionary, another input mode is required, such as Multitap.

4.1.3 Two-Key

The Two-Key method is not popular, but it exemplifies the diverse possibilities of the 12-key keypad. It involves first pressing a key to choose a character group, then pressing 1, 2, 3, or 4 to select a character within the group.

4.1.4 Less-Tap

Less-Tap is a variation on Multitap wherein the letters on each key are ordered according to their frequency in the language. For English, the 2 key cycles the characters in order acb rather than abc, because c appears more often than b (Pavlovych & Stuerzlinger, 2004). Although fewer keystrokes are required, attention demands are initially higher since the user must attend to the display to observe the effect of each key press.

4.1.5 TiltText

In TiltText (Wigdor & Balakrishnan, 2003), tilting the phone in one of four directions results in the selection of a character associated with the direction and key pressed. For example, to obtain an a, the phone is tilted to the left and 2 is pressed. A similar procedure while tilting right produces c, etc. TiltText results in faster input than the MultiTap method.

4.1.6 LetterWise

LetterWise (Eatoni Ergonomics, New York) uses prefix-based disambiguation utilizing a database of letter sequences and their probabilities of appearing in the language. The most-probable letter is presented based on previously entered characters. If the presented letter is not correct, the user presses the “next key” until the desired letter appears. An advantage over T9 is that both dictionary and non-dictionary words are entered with similar success. However, attention demands are high, since the result of each key press must be monitored.

4.2 Keypad Types

There are two main types of 12-key keypads in use. The traditional keypad employs physical buttons with tactile and aural feedback. A newer method is virtual keypads, which display simulated buttons on a small screen. Virtual keypads are found where size constraints preclude physical keypads. The methods above can be implemented on either type of keypad. However, touch-sensitive displays generally lack innate tactile and aural feedback which adversely affects speed and throughput (MacKenzie & Oniszczak, 1998). To address this, adding hardware such as a piezzo device under a touch screen have been successfully demonstrated (Poupyrev, Maruyama, & Rekimoto, 2002). Some recent phones such as the LG LX-570 Muziq[12] make use of the vibrate feature found in most phones (which is accomplished via a small motor which has an off-center weight attached to the spindle to cause a vibrating effect when activated) to vibrate as one touches its smooth surface controls. The addition of a mechanical relay to a keypad has not yet been explored. However, there are potential advantages to doing so. A “click” of a mechanical relay more realistically simulates the sound of a physical button being pressed as well as convincingly emulating the feel of a button being clicked by providing a small localized “kick”. Additionally, one may simulate the hysteresis of a physical button springing back by deactivating the relay at a reduced pressure level which results in an “unclick” sound and feel. The RollPad method described below makes use of just such a system.

4.3 RollPad

RollPad is a new method of inputting text onto a 12-key soft keypad. While soft keypads are generally implemented as a compromise, RollPad leverages a property of finger-activated touchscreens and touchpads that physical buttons lack – the ability to be “rolled”. For example, to input the letter f, a press of the 3 key with a slight rolling motion of the finger to the right produces the desired character. The direction of the roll corresponds to the position of the letter in the triad. So, pressing 3 with a left roll produces d, pressing 3 with no roll produces the middle e, and pressing 3 with a right roll produces the rightmost character, f. Keys 2–9 operate in this way while key 1, produces q with a leftward roll and a z with a rightward roll. See Figure 14.

Figure 14: The RollPad in Use

Some implementations of soft keyboards involve a steep learning curve (MacKenzie & Zhang, 1999). To further improve speed and accuracy of the device and minimize additional learning, aural and tactile feedback was added as was done with the earlier Tactile Touchpad (MacKenzie & Oniszczak, 1998).

5 Experiment 1 – Tactile Touchpad

The first experiment involves a Fitts’ Law task to measure speed, accuracy and throughput. It compares the new Tactile Touchpad to traditional Lift-and-Tap and button selection techniques for touchpads.

5.1 Method

5.1.1 Participants

Twelve participants (5 male, 7 female) were used in the study. All participants were right handed, and all used computers with graphical user interfaces on a daily basis. Two participants had prior experience with touchpads.

5.1.2 Apparatus

An Intel Pentium-class system with a 17" monitor set at a resolution of 1024 x 768 pixels was used. The Ctmouse mouse driver for the MS-DOS operating system, version 1.2, was used for all but the Tactile Touchpad condition. For the latter, a custom driver was written to implement the special features of the tactile condition.

The experiment used custom software known as the Generalized Fitts’ Law Model Builder (W. Soukoreff & MacKenzie, 1995). The software executes under MS-DOS 5.0 and interacts with the system’s pointing device through the installed mouse driver.

All three selection techniques used the same device, a modified Synaptics T1002D touchpad, as described earlier. Standard features of the touchpad include two physical buttons and a lift-and-tap button emulation in firmware.

For each block of trials the experimental software presented a new target condition. Two rectangles of width W separated by distance D appeared. The height of the rectangles was set to as tall as possible so that no up or down adjusting of the cursor was necessary to select the targets in the operational parameters of the experiment. A crosshair pointer appeared in the left rectangle and a red X appeared in the opposite rectangle denoting it as the current target (see Figure 15.)

Figure 15: Experimental condition.

5.1.3 Procedure

Participants were instructed to move the pointer by moving their index finger on the touchpad surface. Specifically, they were instructed to move the pointer as quickly and accurately as possible from side to side alternately selecting the target using the current selection technique.

As each target was selected the red X disappeared and reappeared in the opposite rectangle. This helped synchronize participants though a block of trials. If a select operation occurred while the pointer was outside the target, a beep was heard to signal an error. Participants were instructed to continue without trying to correct errors. For each task condition, participants performed 20 selections.

Before gathering data, the task and the selection technique were explained and demonstrated to the participants. Also, participants were given a block of warm-up trials prior to data collection.

Figure 16: Tactile Touchpad in use

5.1.4 Design

The experiment was a 3 × 3 × 3 × 3 × 20 within subjects design. The independent variables were as follows:

Selection Technique	button, lift-and-tap, tactile
Block	1, 2, 3
Target Distance	40, 80, 160 pixels
Target Width	10, 20, 40 pixels
Trial	1, 2, 3 ... 20

The conditions above combined with 12 participants represent a total of 19,440 trials. To minimize skill transfer, the presentation of the selection techniques was counter balanced. The target distance/size conditions were blocked. Each block consisted of nine distance/size combinations presented in random order. For each condition, participants performed 20 trials in succession.

The distance/size conditions were chosen to create a set of tasks covering a range of task difficulties. The easiest task combines the largest target (40 pixels) with the shortest distance (40 pixels). The index of task difficulty is

Equation 4: The equation for index of difficulty with W=40 and D=40

The most difficult task combines the smallest target (10 pixels) with the largest distance (160 pixels):

Equation 5: The equation for index of difficulty with W=10 and D=160

Rest intervals were permitted between blocks of trials. The duration of rest intervals was based on participants’ discretion. All three selection techniques were tested in a single session lasting about an hour. At the end, participants were given a brief questionnaire on their impressions of the three selection techniques.

5.2 Results and Discussion

Since the experiment employed a within-subjects design, a Latin Square was used to balance potential learning effects. However, there remained the possibility of asymmetrical skill transfer (Martin, 2004) from one selection technique to the next based on the order of presentation. This was tested for and was found not to have occurred, as the effect for order of presentation was not statistically significant on all three dependent measures (movement time, error rate, throughput, F_2,9 < 1).

5.2.1 Movement Time

The grand means on the three primary dependent measures were 1641 ms for movement time, 6.6% for error rate, and 1.17 bps for throughput. The interaction technique and block effects on these measures are reported in the following sections.

5.2.2 Speed and Accuracy

The tactile selection technique had the lowest movement time per trial at 1345 ms. The other conditions were slower by 20% for lift-and-tap (1611 ms) and by 46% using the physical button (1967 ms). These differences were statistically significant (F_2,18 = 47.6, p < .0001).

Exactly the opposite ranking was observed on error rates, however. Using a button for the select operation, the error rate was 4.1%. It was 1.4 times higher using lift-and-tap (5.8%) and 2.4× higher using the tactile condition (9.9%). However, these differences were not statistically significant (F_2,18 = 2.27, p > .05).

The results for speed and accuracy are shown in Figure 17. Overall performance is better toward to bottom-left of the figure.

Figure 17: Results for speed and accuracy

5.2.3 Throughput

A strong analysis of the effect of selection technique is obtained by the dependent measure throughput, because it reflects both the speed and accuracy of performance and because it is the measure recommended in the ISO standard, 9241-9. The highest throughput was observed in the tactile condition at 1.43 bps. The other conditions exhibited lower throughputs by 25% for lift-and-tap (1.07 bps) and by 31% using a button (0.99 bps). See Figure 18 for details. The differences were statistically significant (F_2,18 = 18.0, p < .0001).

Figure 18: Throughput by selection technique

These measures for throughput are on the low end of the range as compared to other pointing devices. In unpublished studies using the same experimental conditions, measures in the range of 3.0-4.5 bps for mice and 2.0-3.5 bps for trackballs were found. Published figures for throughput are also higher, in general. A 1991 study reported 3.3 bps for a Kensington trackball, 4.5 bps for an Apple mouse, and 4.9 bps for a Wacom stylus (MacKenzie et al., 1991), while a 1993 study found throughput equal to 4.3 bps for the mouse (MacKenzie & Ware, 1993). Rates less than 4 bps are not uncommon, however (Balakrishnan, Baudel, Kurtenbach, & Fitzmaurice, 1997; Boritz, Booth, & Cowan, 1991; Gillan, Holden, Adam, Rudisill, & Magee, 1990; Gillan, Holden, Adams, Rudisill, & Magee, 1992; MacKenzie, 1995).

(a)

(b)

(c)

Figure 19: Block by interaction technique for (a) movement
time, (b) error rate, and (c) throughput.

5.2.4 Learning Effects

For each selection technique, participants performed three blocks of trials in succession. Each block consisted of 20 trials on each of the nine randomly presented target conditions (180 total trials). It is worthwhile, therefore, to examine the effect of “block” on the three dependent measures, since this reflects the extent to which participants improved with practice. As well, a block x selection technique interaction effect may be present, indicating different learning patterns across devices.

The main effect of block was statistically significant for movement time and throughput, but not for error rate. The reverse pattern emerged for the block × selection technique interaction, which was significant for error rate, but not for movement time or throughput. These patterns are illustrated in Figure 19.

The pattern in all three parts of Figure 19 looks favorable for the tactile selection condition. The improvement in performance is clearly seen in each figure, and it is most dramatic from block 2 to block 3 (although the block × interaction technique effect was not statistically significant). With continued practice, the tactile condition is likely to improve. On error rate - the only measure on which the tactile condition faired poorly - it might even “catch up”, although this could only be determined in a prolonged study.

5.2.5 Outliers

Since the error rates were somewhat high, further investigation was warranted. A category of response called “wrong-side outliers” was identified. These were selections that occurred on the wrong side of the display. For example, if the goal was to select the target on one side of the display and the selection occurred before the pointer was halfway to the target the selection was on the wrong side of the display. This is a gross error. These are considered “outliers” because they are outside the normal range of variations expected in participants’ behavior. A wrong-side outlier can occur for several reasons, such as double-clicking on a target or inadvertent lifting or pressing with the finger during pointer motion.

Overall, button selection had the fewest wrong-side outliers (178, 2.75%), followed by tactile (245, 3.78%) and lift-and-tap (253, 3.90%). Comparing the percentages with the overall error rates given earlier, it is seen that wrong-side outliers formed a significant portion of the overall errors.

The number of wrong-side outliers, by selection technique and block, is shown in Figure 20.

Figure 20: Wrong-side outliers by block and selection technique

The good showing of the button technique is likely due to the clear separation of pointer movement from target selection. Since movement and selection are more integrated with the lift-and-tap and tactile conditions, higher rates for wrong-side outliers are expected.

5.2.6 Questionnaire

At the end of the experiment, participants were given a questionnaire. For each selection technique, they were asked to provide a rating on their speed perception, their accuracy perception, and their overall preference. They entered one of four possible values from 1 (slowest, least accurate, liked the least) to 4 (quickest, most accurate, liked the most). The results are shown in Table 2. Each cell is the total score for twelve participants, with higher scores preferred.

Table 2: Questionnaire results. (Note: Scores are totals of
participants’ ratings; higher scores are better.)

Selection Technique	Speed Perception	Accuracy Perception	Overall Preference
Button	3	15	5
Lift-and-tap	15	13	15
Tactile	19	15	17

Participants liked the tactile selection technique. (This was evident in their comments, as well.) Tactile selection ranked 1st for speed perception, 1st (tied) for accuracy perception, and 1st for overall preference. It is noteworthy that on accuracy participants rated the tactile condition equal to, or better than, the other conditions even though it had the highest error rate. This could be due to the higher measures for throughput, which reflect the overall ability of participants to complete their tasks.

6 The Potential for a Tactile Touchpad

The Synaptics touchpad’s method of deriving pressure data is indirect since it senses the capacitance between the finger and the pad. Pressure is derived from the area of the user’s finger contacting the surface of the pad. Since one’s finger flattens on the pad with increased pressure, the device takes advantage of this correlation. As a consequence, users with small fingers must press harder than users with large fingers. Participants with particularly large fingers required a more delicate touch than they preferred. This may account for the increased error rate of the Tactile Touchpad condition.

A better version of the touchpad would use the true pressure-sensing technology of force sensing resistor touchpads such as the MicroNav Pad (formerly VersaPad) by Interlink Electronics[13]. A future replication of this experiment utilizing a calibration procedure at the onset would also be interesting, although this is generally not considered acceptable as a required procedure in commercial pointing devices.

Another noticeable artifact of the Tactile Touchpad condition was a tendency for the on-screen pointer to move down slightly as the participant pressed down to select a target. This was most pronounced with participants who held their pointing finger relatively perpendicular to the touchpad’s surface. When they pressed down, the center of the finger’s surface area moved towards the bottom and the onscreen pointer momentarily shifted downward with each press. As the targets were long and vertical, this most likely did not have an effect in the experiment; however, it is noteworthy. One participant suggested that the pointer freeze at a certain pressure level prior to a button press registering so that the results would be more predictable. Another possible solution would be to correct for the downward dips as the user pressed on the pad through software. That is, as the “pressure” increased, the pointer’s vertical value might be slightly increased to compensate for the user’s tendency to move the pointer downwards.

6.1 Summary

Although touchpads are not likely to supplant mice on the desktop, the results have implications for portable computer usage, and further refinements may make the Tactile Touchpad closer to a mouse in performance.

The Tactile Touchpad was found superior to both the lift-and-tap mode touchpad and button mode touchpad in terms of movement time and throughput. Although the error rate was higher than with the other touchpad conditions, it was not generally noticed by the participants and the overall flow of information (viz., throughput) was higher even with the increased error rate. With design improvements, the use of embedded tactile feedback in a touchpad can facilitate simple interactions such as pointing and selecting.

7 Experiment 2 - UniPad

The second experiment involved evaluating the UniPad technique of text entry with a method containing no language-based acceleration text entry aids.

7.1 Method

7.1.1 Participants

Ten paid volunteer participants (6 male, 4 female) were recruited from the local university campus. Participants ranged from 18 to 35 years (mean = 26.1, sd = 5.9). All were daily users of computers, reporting 3 to 8 hours usage per day (mean = 5.6, sd = 2.2). Self-assessed typing speeds ranged from 35 to 52 words per minutes (mean = 42.2, sd = 4.7). Two indicated prior experience with Graffiti, describing their expertise as “novice” or “intermediate”.

7.1.2 Apparatus

The experiment was conducted in a research lab using a standard desktop computer system equipped with a 13.3” diagonal Wacom PL-400 tablet for stylus entry. The PL-400 tablet is both a digitizer for input and a 1024 x 768 LCD colour screen for output.

The experimental software was a custom Java application for text entry evaluation. The candidate list appeared at the first stroke. Other settings were as follows:

- Candidate list size: 5

- Frequent word prompt list size: 12

- Suffix list size: 12

The experiment uses a set of 500 phrases ranging from 16 to 43 characters (mean = 28.6) (MacKenzie & Soukoreff, 2003). There were 2712 total words, including 1163 unique words. Words ranged from 1 to 13 characters (mean = 4.46). The correlation between the letter frequencies in the phrase set and those in the reference corpus was r = .9541. During execution, phrases are selected randomly and presented to the participant for input.

7.1.3 Procedure

Participants completed a pre-test questionnaire soliciting demographic and computer usage information (results cited above) and a post-test questionnaire on their subjective impressions of the methods (discussed later).

Each participant completed two one-hour sessions. Plain single stroke text entry was used in the first session. This served to bring participants up to speed on learning the single stroke alphabet (see Figure 7 on page 19). In the second session, word-completion was introduced.

Prior to collecting data, the experimenter briefly explained the task and demonstrated the software including the method to correct errors and to terminate entry at the end of a phrase. The single-stroke alphabet was displayed on a wall chart. Acceleration features were explained and demonstrated before the second session.

The participants were instructed to enter the phrases “as quickly and accurately as possible”. The task was described as “similar to sending an email message to a friend” – the message should be understandable, but not necessarily perfect.

The goal of the instructions was to elicit typical text input behaviors and therefore improve the external validity of the experiment. However, a side effect is that there are two types of errors: those corrected, and those remaining in the transcribed text. The former are reflected in the measured KSPC as additional strokes to backup and correct mistakes. The latter are measured by comparing the presented and transcribed text phrases and computing the error rate. The error rate was computed using the minimum string distance method (R. W. Soukoreff & MacKenzie, 2001).

Participants entered a few warm-up phrases during which time they could ask questions about the procedure. Data collection began with the first stroke for each phrase and ended with a tap on the ENTER soft-key. Participants were allowed to rest at their discretion between phrases.

Each session was divided into 4 blocks of 15-minutes each. Hence blocks 1–4 refer to data from session one (plain single stroke entry method), while blocks 5-8 are for session two (single stroke method with word completion). Participants were encouraged to take a short break between blocks.

The software recorded a timestamp for each stroke, as well as a variety of statistics for follow-up analyses.

7.2 RESULTS AND DISCUSSION

In all, participants entered 754 phrases using the single-stroke method (blocks 1–4) and 937 phrases using the single-stroke method with word completion, suffix completion, and frequent word prompting (blocks 5–8).

7.2.1 Outliers

In word completion mode, an anomalous behavior was observed. The Enter soft key, to terminate entry of a phrase, was occasionally tapped to select a word. This prematurely terminated the phrase, resulting in an abnormally high error rate. These trials were classified as outliers and were removed. In all, 30 of the 937 phrases (3.2%) were eliminated by applying this criterion.

7.2.2 Speed, Accuracy, and KSPC

The results for speed are shown in Figure 21a. The first block rate of 6.6 wpm is close to the 7 wpm novice rate for Graffiti reported by Fleetwood et al (2002). As expected, entry speed increased significantly with practice (F_1,9= 40.7, p < .0001). There was also a significant increase in speed by entry mode (F_3,27= 58.1, p < .0001). The average entry rate for block 8 (single stroke method with word completion) was 12.8 wpm, with rates for individual participants ranging from 8.0 to 18.2 wpm. The block ´ entry mode interaction was not significant, however (F_3,27= 0.45, ns).

Averaged by block, error rates were below 3% throughout the experiment. As seen in Figure 21b, error rates improved during blocks 5–8, dropping to just under 1% by the end of the experiment. The lower error rates during blocks 5–8 are likely due to a cognitive shift to “word-level attention” when using word completion. More accurate entry is also a likely by-product of the reduced KSPC when using word completion. Participants varied quite a bit in their disposition for accuracy. The main effect was not statistically significant for block (F_1,9= 2.7, p > .05) or entry mode (F_3,27= 0.54, ns).

(a)

(b)

(c)

Figure 21: Results for the single stroke method (blocks 1–4) and the single stroke method with word completion (blocks 5–8). (a) speed (b) error rate (c) keystrokes per character

Figure 21c clearly shows two trends for KSPC. During blocks 1–4, KSPC was relatively stable, with means of 1.417 in block 1 and 1.301 in block 4, representing a drop of 8.2%. While using language-based acceleration (blocks 5-8), KSPC was markedly less – and below 1.000, as conjectured. An improvement of 6.0% occurred with KSPC dropping from 0.793 in block 5 to 0.746 in block 8. These improvements were statistically significant, as demonstrate in the F-statistics for entry mode (F_1,9= 857.3, p < .0001) and block (F_3,27= 9.40, p < .0005). The block ´ entry mode interaction was not significant, however (F_3,27= 1.44, p > .05).

Errors followed by corrective actions tend to push the observed KSPC up, as do unrecognized strokes, or characters erroneously inserted and left in the transcribed text. During blocks 5–8, KSPC is also pushed up by participants failing to capitalize on opportunities, such as selecting a word when it first appears in the candidate list.

The most significant observation is the failure of UniPad’s language-based acceleration features to yield a significant increase in text entry throughput – an increase consistent with the decrease in keystrokes afforded by word completion and the related acceleration features. The average observed KSPC for block 8 (0.746) is still (0.746–0.428) / 0.428 = 74.2% above the optimal value.

7.2.3 Participant Questionnaire

The post-test questionnaire (Table 3) solicited comments and responses to seven statements. The results are shown below. (Note: strongly disagree = 1; disagr