This paper describes two strategies, operating at different levels of speech, which exploit special characteristics of the speech understanding task; both involve word islands within an utterance. First, a new upper bound based on the probability of the best possible parse is proposed for scoring partial interpretations of an acoustic signal. Subsequently, the paper describes a method of automating rule discovery for semantic parsers. The rules are incorporated in a structure called a String Classification Tree and involve patterns of key words; they are robust in the presence of production and recognition errors.