Pickup Usability Dominates - Human-Computer Interaction and Innovation in Handheld, Mobile and Wearable Technologies - page 47

disambiguation; here users use an alternative
input method, e.g. tilting the phone (Wigdor &
Balakrishnan, 2003) or a small chord-keyboard
on the rear of the phone (Wigdor & Balakrishnan,
2004), to disambiguate the letter as it is entered.
While clearly potentially much faster than multi-
tap and relatively easy to use, this approach has
not yet been picked up by device manufacturers.
Aimed at overcoming the problems of multi-
tap, predictive text entry approaches use language
modelling to map from ambiguous codes to words
so that users need only press each key once, for
example mapping the key sequence 4663 directly
to good . While there are clearly cases where there
are more than one match to the numeric key se-
quence (e.g. 4663 also maps to home and gone
amongst others), these are surprisingly rare for
common words. The problem of multiple matches
can be alleviated to a large extent by giving the
most likely word as the first suggestion then al-
lowing users to scroll through alternatives for less
likely words. Based on a dictionary of words and
their frequency of use in the language, users get
the right word suggested first around 95% of the
time (Gong & Tarasewich, 2005). AOL-Tegic's
T9 (Grover, King, & Kushler, 1998; Kushler,
1998) industry-standard entry method is based
around this approach and is now deployed on
over 2 billion handsets. Controlled experiments
have shown this form of text entry considerably
out-performs multi-tap (Dunlop & Crossan, 2000;
James & Reischel, 2001), with text entry rising
from around 8wpm for multitap to around 20 for
T9. While predictive text entry is very high qual-
ity, it is not perfect and can lead to superficially
unrelated predictions that are undetected as users
tend to type without monitoring the screen (e.g.
a classic T9 error is sending the message call
me when you are good rather than are home ).
The main problem, however, with any word
prediction system is handling out-of-vocabulary
words—words that are not known to the dictionary
cannot be entered using this form of text entry.
The usual solution is to force users into an “add

word” dialogue where the new word is entered
in a special window using multi-tap—clearly at
a considerable loss of flow to their interaction
and reduction in entry speed. As most people do
not frequently enter new words or place/people
names, this is not a major long-term problem.
However, it does considerably impact on initial
use and can put users off predictive text messaging
as they constantly have to teach new words to the
dictionary in the early days of using a new device.
This in turn impacts on consumer adoption with
many people not using predictive text despite it
clearly being faster for experienced users.
An alternative approach to dictionary and
word-level disambiguation is to use letter-by-letter
disambiguation where letters are suggested based
on their likelihood given letters already entered
in the given word or likely letters at the start of
a word (e.g. in the clearest case in English, a q
is most likely to be followed by a u ). This gives
the user freedom to enter words that are not in the
dictionary and considerably reduces the memory
load of the text entry system (no longer an issue
with phones but still an issue on some devices).
Experiments using this approach (MacKenzie,
Kober, Smith, Jones, & Skepner, 2001) showed
key-strokes halved and speed increased by around
36% compared with multi-tap. They also claim
that this speed is inline with T9 entry and that
their approach out-performs T9 by around 30%
when as few as 15% of the least common words
are missing from the predictive dictionary. Predict-
ing letters based on previous letters is actually a
specific implementation of Shannon's approach to
prediction based on n-grams of letters (Shannon,
1951). Some work has been carried out to extend
this to the word-level and shows good promise: for
example bi-gram word prediction in Swedish with
word completion reduced keystrokes by between
7 and 13% when compared with T9 (Hasselgren,
Montnemery, Nugues, & Svensson, 2003).
In work on watch-top text entry we (Dunlop,
2004) found that moving to a 5-key pad reduced
accuracy from around 96% to around 81% with