H94-1079 |
parameters M new using 1 iteration of
|
Viterbi training
|
. 6 . Update the model by smoothing
|
E09-1020 |
number of co-occurrences . A greedy
|
Viterbi training
|
is then applied to improve this
|
J03-1002 |
simplest method is to perform
|
Viterbi training
|
using only the best alignment
|
C96-1051 |
quantization and beam-search driven
|
Viterbi training
|
and recognition . The ls ` adora
|
J04-1004 |
Su ( 1997 ) use an unsupervised
|
Viterbi training
|
process to select potential unknown
|
J12-3003 |
( 2010c ) for the hardness of
|
Viterbi training
|
and maximizing log-likelihood
|
D13-1204 |
single ( unlexicalized ) step of
|
Viterbi training
|
: The idea here is to focus on
|
J93-2003 |
Viterbi alignment , we call this
|
Viterbi training
|
. It is easy to see that Viterbi
|
J93-2003 |
lies in better modeling . 6.2
|
Viterbi Training
|
As we progress from Model 1 to
|
D13-1204 |
the " baby steps " strategy with
|
Viterbi training
|
( Spitkovsky et al. , 2010 ,
|
H93-1020 |
training is similar to IIMM "
|
Viterbi training
|
" , in which training data is
|
J93-2003 |
then a similarly reinterpreted
|
Viterbi training
|
algorithm still converges . We
|
D11-1117 |
settings that are least favorable for
|
Viterbi training
|
: adhoc and sweet on . Although
|
D11-1117 |
average , compared to standard
|
Viterbi training
|
; A43 is , in fact , 20 % faster
|
D11-1117 |
- erage , compared to standard
|
Viterbi training
|
; A13 is only 30 % slower than
|
D11-1117 |
set-up may be disadvantageous for
|
Viterbi training
|
, since half the settings use
|
D13-1204 |
also as a single unlexicalized
|
Viterbi training
|
step , but now with proposed
|
H91-1052 |
Since these both were essentially
|
Viterbi training
|
procedures ( estimated from only
|
D13-1204 |
( single steps of lexicalized
|
Viterbi training
|
on clean , simple data ) , ahead
|
D11-1117 |
average , compared to standard
|
Viterbi training
|
; A23 is again 30 % slower than
|