The performance of a spoken language processing system depends heavily on the ambiguity resolution capability of the underlying language model. Conventional stochastic context-free grammars do not provide contextual information and semantic constraints required for ambiguity resolution. In our formulation, a syntactic score function could be used to enhance the disambiguation capability by taking context-sensitivity into account; a semantic score could be used to further help resolve ambiguities in syntactic and semantic levels. Furthermore, a baseline model may not perform satisfactorily due to unreliable estimation of the parameters or lack of discrimination power and robustness of the baseline model. Therefore, it is desirable to increase the discrimination power and robustness of the baseline system. In this paper, an adaptive learning algorithm, with discrimination and robustness emphasized, is used to improve the baseline systems for ambiguity resolution. Performance improvement over the baseline system will be shown for some preliminary experiments.