Date post: | 24-May-2015 |
Category: |
Science |
Upload: | hugo-nicolau |
View: | 202 times |
Download: | 15 times |
B# :: Chord-based Correction for Multitouch Braille InputHUGO NICOLAUKYLE MONTAGUETIAGO GUERREIROJOÃO GUERREIROVICKI L. HANSON
motivation :: non-visual input
motivation :: braille input
[Southern et al. 2012]
[Oliveira et al. 2011] [Azenkot et al. 2012]
motivation :: fast typing (4x)
problem :: error prone
50%
challenge :: accurate input
approach :: spellchecker
? ? ? ? ? ? ? ? ? ?
> e d k l o w u h k d _
approach :: character-level
e x i l e d u n k s
> e d k l o w u h k d _
B# :: chord-level spellchecker
h e l l o w o r l d
> e d k l o w u h k d _
B# :: data collection
B# :: data collection
11 blind participants :: ages 22 – 62 ::
3 females :: grade 1 braille ::
no touchscreen experience
major results :: substitutions
Insertions Substitutions Omissions
23%
72%
5%
Overall Error Rate: 18%
major results :: 43% single-dot error
omission
insertion
transposition
q g
i f
b k
major results :: invalid characters
! t
B# :: braille distance
[100010]
[100010]
= ?
-
distance :: damerau-levenshtein
[100] – [110] + [010] – [010]
= 1 + 0 = 1
B# :: scoring function
scoreword = wMSD(Chtyped, Chword) + ƒword
evaluation :: B#, AOSP
evaluation :: dataset
751 words :: 49% incorrect entries ::
word list with 213,133 entries
results :: B# is more accurate
1 2 3 4 5 6 7 8 9 1025%
35%
45%
55%
65%
75%
56%
75%
28%39%
Acc
ura
cy
B#
Android
results :: word correction
204 of 364
100 of 364
results :: no gain beyond 5
1 2 3 4 5 6 7 8 9 1025%
35%
45%
55%
65%
75%A
ccura
cy
B#
Android
6%
9%
results :: incorrectly corrected
Android B#
1.4%1.7%
conclusion :: B#
conclusion :: higher accuracy
2X
future :: non-visual interface
The End.HUGO [email protected], slides, and video @ http://web.ist.utl.pt/hugo.nicolau
non-visual correction :: challenges
Automatic correction?
Word-completion?
How to provide corrections?
When to provide them?
How to users explore suggestions?
B#: future improvements
Bi-grams, Tri-grams, N-grams
“Machine Learned” Distance Function
Personalized spellchecker
Leverage touch information
B#: android’s proximity matrix
Fairly complex spellchecker
Dozen of features
Worthwhile doing