model,11-1-H01-1058,bq address the problem of combining several <term> language models ( LMs ) </term> . We find that simple <term> interpolation
tech,4-2-H01-1058,bq LMs ) </term> . We find that simple <term> interpolation methods </term> , like <term> log-linear and linear
tech,8-2-H01-1058,bq interpolation methods </term> , like <term> log-linear and linear interpolation </term> , improve the <term> performance </term>
measure(ment),15-2-H01-1058,bq interpolation </term> , improve the <term> performance </term> but fall short of the <term> performance
measure(ment),21-2-H01-1058,bq performance </term> but fall short of the <term> performance </term> of an <term> oracle </term> . The <term>
other,24-2-H01-1058,bq of the <term> performance </term> of an <term> oracle </term> . The <term> oracle </term> knows the
other,1-3-H01-1058,bq </term> of an <term> oracle </term> . The <term> oracle </term> knows the <term> reference word string
other,4-3-H01-1058,bq </term> . The <term> oracle </term> knows the <term> reference word string </term> and selects the <term> word string </term>
other,10-3-H01-1058,bq word string </term> and selects the <term> word string </term> with the best <term> performance </term>
measure(ment),15-3-H01-1058,bq <term> word string </term> with the best <term> performance </term> ( typically , <term> word or semantic
measure(ment),19-3-H01-1058,bq <term> performance </term> ( typically , <term> word or semantic error rate </term> ) from a list of <term> word strings
other,29-3-H01-1058,bq error rate </term> ) from a list of <term> word strings </term> , where each <term> word string </term>
other,34-3-H01-1058,bq <term> word strings </term> , where each <term> word string </term> has been obtained by using a different
model,43-3-H01-1058,bq been obtained by using a different <term> LM </term> . Actually , the <term> oracle </term>
other,3-4-H01-1058,bq different <term> LM </term> . Actually , the <term> oracle </term> acts like a <term> dynamic combiner
tech,7-4-H01-1058,bq the <term> oracle </term> acts like a <term> dynamic combiner </term> with <term> hard decisions </term> using
other,10-4-H01-1058,bq a <term> dynamic combiner </term> with <term> hard decisions </term> using the <term> reference </term> .
other,14-4-H01-1058,bq <term> hard decisions </term> using the <term> reference </term> . We provide experimental results
tech,11-5-H01-1058,bq results that clearly show the need for a <term> dynamic language model combination </term> to improve the <term> performance </term>
measure(ment),18-5-H01-1058,bq model combination </term> to improve the <term> performance </term> further . We suggest a method that
skrij podrobnosti