Problem in discrminiative models

In dialog act tagging, if we use discriminative model, the links in words won't affect the results. How to solve this problem?