Natural Language Generation for dialogue: system survey

Many natural language dialogue systems make use of `canned text' for output generation. This approach may be su±cient for dialogues in restricted domains where system utterances are short and simple and use fixed expressions (e.g., slot filling dialogues in the ticket reservation or travel information domain); but for more sophisticated dialogues (e.g., tutoring dialogues) a more advanced generation method is required. In such dialogues, the system utterances should be produced in a context-sensitive fashion, for instance by pronominalising anaphoric references, and by using more or less elaborate wording depending on the state of the dialogue, the expertise of the user, etc. In the case of spoken dialogues, it is very useful if the natural language generation component can provide information that is relevant for determining the prosody of the speech output. Similarly, for use in embodied agents it is useful if the generation component can provide information about the facial and body movements that should accompany the language being produced by the agent. Clearly, it will be extremely di±cult to achieve all this using simple string manipulation, so a more flexible and context-sensitive generation method is required. This report discusses some of the possibilities for the sophisticated generation of system utterances in a dialogue system. The basic assumption is that this task is performed by a separate language generation module, which takes as its input a message specification produced by a dialogue planner and transforms this message into an expression in natural language. Part I of this report provides a general discussion of different methods for performing this task, and outlines some requirements on language generation systems that might be used for this purpose. Part II gives an overview of publicly available language generation systems, and discusses to what extent they meet the previously stated requirements