|Brian McMahan, Matthew Stone|
An appealing methodology for natural language generation in dialogue systems is to train the system to match a target corpus. We show how users can provide such a corpus as a natural side effect of interacting with a prototype system, when the system uses mixed-initiative interaction and a reversible architecture to cover a domain familiar to users. We experiment with integrated problems of sentence planning and realization in a referential communication task. Our model learns general and context-sensitive patterns to choose descriptive content, vocabulary, syntax and function words, and improves string match with user utterances to 85.8% from a handcrafted baseline of 54.4%.