Automatic identification of command boundaries in a conversational natural language user interface

In this paper, we propose a trainable system that can automatically identify the command boundaries in a conversational natural language user interface. The proposed solution makes the conversational interface much more user friendly, and allows the user to speak naturally and continously in a hands-free manner. The main ingredient of the system is the maximum entropy identification model, which is trained using data that has all the correct command boundaries marked. During training, a set of features and their weights are selected iteratively using the training data. The features consists of words and phrases, as well as their relative position to the potential command boundaries. Decoding is done by examining the product of the weights for the features that are present. We also propose several enhancements to the approach, such as combining it with a more effective language model at the speech recognition stage to generate additional tokens for the identification model. We conducted several experiments to evaluate the proposed approach, and the results are described.