Abstract

The semantic frame based spoken language understanding involves
two decisions – frame classification and slot filling. The
two decisions can be made either separately or jointly. This
paper compares the different strategies and presents some empirical
results in the conditional model framework when only a
small amount of training data is available. It is found that while
the two pass classification/slot filling solution has resulted in
the much better frame classification accuracy, the joint model
has yielded better results for slot filling. Application developers
need to carefully choose the strategy appropriate to the application
scenarios.