Semantics-oriented language understanding with automatic adaptability

We propose a novel language understanding approach, the composition of classification, for utterances interpretation in a dialogue system. All the classification results including both the cascading information and the meaning information are composed to a nested semantic frame. This bypasses the traditional syntactic parsing and extracts semantics directly from plain text. To overcome the problem of data sparsity of dialogue systems and ease the annotation of the training data, we are the first to report to generate and annotate the training data using a finite state machine.