#3075In this paper, we show how training data can be supplemented with text from the web filtered to match the style and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from the data by usingclass-dependent interpolation of N-grams.
other,9-1-N03-2003,ak
for
<term>
language modeling
</term>
of
<term>
conversational speech
</term>
are limited . In this paper , we
#3024Sources of training data suitable for language modeling ofconversational speech are limited.
lr,43-2-N03-2003,ak
bigger performance gains from the
<term>
data
</term>
by using
<term>
class-dependent interpolation
#3072In this paper, we show how training data can be supplemented with text from the web filtered to match the style and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from thedata by using class-dependent interpolation of N-grams.
tech,6-1-N03-2003,ak
<term>
training data
</term>
suitable for
<term>
language modeling
</term>
of
<term>
conversational speech
</term>
#3021Sources of training data suitable forlanguage modeling of conversational speech are limited.
other,49-2-N03-2003,ak
class-dependent interpolation
</term>
of
<term>
N-grams
</term>
. In order to boost the
<term>
translation
#3078In this paper, we show how training data can be supplemented with text from the web filtered to match the style and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation ofN-grams.
other,21-2-N03-2003,ak
<term>
web
</term>
filtered to match the
<term>
style
</term>
and/or
<term>
topic
</term>
of the
<term>
#3050In this paper, we show how training data can be supplemented with text from the web filtered to match thestyle and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation of N-grams.
tech,26-2-N03-2003,ak
</term>
and/or
<term>
topic
</term>
of the
<term>
target recognition task
</term>
, but also that it is possible to
#3055In this paper, we show how training data can be supplemented with text from the web filtered to match the style and/or topic of thetarget recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation of N-grams.
other,13-2-N03-2003,ak
data
</term>
can be supplemented with
<term>
text
</term>
from the
<term>
web
</term>
filtered
#3042In this paper, we show how training data can be supplemented withtext from the web filtered to match the style and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation of N-grams.
other,23-2-N03-2003,ak
match the
<term>
style
</term>
and/or
<term>
topic
</term>
of the
<term>
target recognition task
#3052In this paper, we show how training data can be supplemented with text from the web filtered to match the style and/ortopic of the target recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation of N-grams.
lr,7-2-N03-2003,ak
limited . In this paper , we show how
<term>
training data
</term>
can be supplemented with
<term>
text
#3036In this paper, we show howtraining data can be supplemented with text from the web filtered to match the style and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation of N-grams.
lr,2-1-N03-2003,ak
tagging result
</term>
. Sources of
<term>
training data
</term>
suitable for
<term>
language modeling
#3017Sources oftraining data suitable for language modeling of conversational speech are limited.
other,16-2-N03-2003,ak
supplemented with
<term>
text
</term>
from the
<term>
web
</term>
filtered to match the
<term>
style
</term>
#3045In this paper, we show how training data can be supplemented with text from theweb filtered to match the style and/or topic of the target recognition task, but also that it is possible to get bigger performance gains from the data by using class-dependent interpolation of N-grams.