 |
ITRW on
|
 |
|
ITRW on
|
|
In this paper we address the problem of mismatch in train and test conditions. Counter intuitive as it may seem, we do this by employing a particular element from the well-known training paradigm of Minimum Classification Error training. Rather than recognising a sentence according to maximum likelihood, we examine a number of likelihood ratio-based word score techniques in order to rescore and resort N-best lists.
Experiments for a Dutch city name recognition task did not lead to improved recognition performance. Analysing the results however, we see a number of promising handles for more succesful attempts in the future. We find cues that the information modelled in antimodels is not only useful for keyword spotting and confidence measure assessment, but may be valuable for decoding as well.
Bibliographic reference. Bouwman, Gies / Boves, Louis (2001): "Using discriminative principles for recognising city names", In Adaptation-2001, 109-112.