Back Room Front Room 2
Back Room Front Room 2
Back Room Front Room 2
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
284 ENTERPRISE INFORMATION SYSTEMS VI<br />
Higginbotham, D. J., Lesher, G. W., and Moulton, B. J.<br />
(1998). Techniques for augmenting scanning communication.<br />
Augmentative and Alternative Communication,<br />
14:81–101.<br />
ISAAC (1983). Int. Soc. for Augmentative and Alternative<br />
Communication, Internet site. http://www.isaaconline.org.<br />
Last accessed October 1th, 2003.<br />
Juang, B. and Rabiner, L. (1990). The segmental k-means<br />
algorithm for estimating parameters of hidden markov<br />
models. In IEEE Trans. on Acoustics Speech and<br />
Signal processing, ASSP-38, pages 1639–1641. IEEE<br />
Computer Society Press.<br />
Juang, B., Rabiner, L., and Wilpon, G. (1986). A segmental<br />
k-means training procedure for connected word recognition.<br />
AT&T Technical Journal, 65:21–31.<br />
Koester, H. and Levine, S. (1994). Learning and performance<br />
of ablebodied individuals using scanning systems<br />
with and without word prediction. Assistive<br />
Technology, page 42.<br />
Lee, H.-Y., Yeh, C.-K., Wu, C.-M., and Tsuang, M.-F.<br />
(2001). Wireless communication for speech impaired<br />
subjects via portable augmentative and alternative system.<br />
In Proc. of the Int. Conf. of the IEEE on Eng.<br />
in Med. and Bio. Soc., volume 4, pages 3777–3779,<br />
Washington, DC, USA.<br />
Prechelt, L. (1996). Early stopping-but when? In Neural<br />
Networks: Tricks of the Trade, pages 55–69.<br />
Quillian, M. (1968). Semantic memory. In Minsky ed.<br />
Semantic Information Processing. MIT Press, Cambridge.<br />
Rabiner, L. (1989). A tutorial on hidden markov models and<br />
selected applications in speech recognition. In Proc. of<br />
the IEEE, 77, pages 257–286. IEEE Computer Society<br />
Press.<br />
Shane, B. (1981). Augmentative Communication: an introduction.<br />
Blackstone, Toronto, Canada.<br />
Simpson, R. C. and Koester, H. H. (1999). Adaptive oneswitch<br />
row-column scanning. IEEE Trans. on Rehabilitation<br />
Engineering, 7:464–473.<br />
Swiffin, A. L., Pickering, J. A., and Newell, A. F. (1987).<br />
Adaptive and predictive techniques in a cammunication<br />
prosthesis. Augmentative and Alternative Communication,<br />
3:181–191.<br />
APPENDIX A<br />
In this appendix, we briefly describe the DAR-HMM according<br />
to the formalism adopted by Rabiner in (Rabiner,<br />
1989) to specify classical hidden Markov models:<br />
• S � {si}, subcategories set with N = |S|;<br />
• V � {vj}, predictable symbols set with M = |V |;<br />
• V (i) = {v (i)<br />
k }, set of symbols predictable in subcategory<br />
i with M (i) = |V (i) | and V = �<br />
i V (i) ;<br />
• O(t) ∈ V , observed symbol at time t;<br />
• Q(t) ∈ S, state at time t;<br />
• πi(t) =P (Q(t) =si), probability that si is the actual<br />
subcategory at time t;<br />
• aii ′ = P (Q(t +1)=si|Q(t) =si ′), transition probability<br />
from si ′ to si;<br />
• b i k = P (O(0) = v (i)<br />
k |Q(0) = si), probability of observing<br />
v (i)<br />
k from subcategory si at t =0;<br />
• b ii′<br />
kk ′ = P (O(t) =v(i)<br />
k |Q(t) =si,O(t − 1) = v(i′ )<br />
k ′ ),<br />
probability of observing v (i)<br />
k from the subcategory si<br />
having just observed v (i′ )<br />
k ′ .<br />
DAR-HMM for symbolic prediction can thus be described<br />
using a parameter vector λ =< Π 0 ,A,B >, where<br />
Π 0 [N] is the vector of inital subcategory probability πi(0),<br />
A[N][N] is the matrix with subcategory transition probabilities<br />
aii ′, and B[N][M][M +1]is the emission matrix with<br />
symbol probabilities b ii′<br />
kk ′ and b i k.Givenλthe vector of parameters<br />
describing a specific language behavior model, we<br />
can predict the first observed symbol as the most probable<br />
one at time t =0:<br />
Ô(0) = arg max<br />
v (i)<br />
k<br />
= argmax<br />
v (i)<br />
k<br />
= argmax<br />
v (i)<br />
k<br />
�<br />
P (O(0) = v (i)<br />
k |λ)<br />
�<br />
(P (O (0) |Q (0) ,λ) P (Q(0)))<br />
�<br />
b i �<br />
k · πi(0) .<br />
Then mimicking the DAR-HMM generative model, to predict<br />
the t th symbol of a sentence we want to maximize the<br />
symbol probability in the present (hidden) state given the<br />
last observed symbol:<br />
�<br />
P O(t) =v (i)<br />
k ,Q(t) =si|O(t − 1) = v(i′ �<br />
)<br />
k ′ ,λ .<br />
Recalling that we can compute the probability of the current<br />
(hidden) state as:<br />
P (Q(t)) =<br />
=<br />
N� P (Q(t)|Q(t − 1)) P (Q(t − 1)) =<br />
N�<br />
i ′ =1<br />
πi ′(t − 1)aii ′ = πi(t),<br />
we obtain a recursive form for symbol prediction at time t:<br />
�<br />
b ii′<br />
N�<br />
�<br />
kk ′ · πi ′(t − 1)aii ′ .<br />
Ô(t) = arg max<br />
v (i)<br />
k<br />
i ′ =1