Within the realization of figure caption, data about the internal state of the visualization as well as data about the interaction history are evaluated in order to generate descriptive figure captions. Natural language generation comprises several stages: text planning, sentence planning and sentence realization. For all these stages several approaches with varying expressive power, flexibility and costs have been suggested in the literature. We will discuss some of them and propose a framework for combining these approaches in order to adapt the generation process to the needs of the graphical application.
The main task of text planning is to decide which portion of the data is to be expressed, considering parameters as the intended user group, user preferences and the interaction history. It provides a representation of the structure of the text to be generated and an explicit or implicit description of the rhetorical structure (the coherence relations between text portions), the intensional structure (which function serves a portion of the text) and the attentional structure .
Within the sentence planning step the text structure is broken down into sequences of sentences. This transformation must determine:
The above mentioned specification from the sentence planning step are put into natural language expression within the sentence realization.