Describing complex charts in natural language: a caption generation system

摘要: Graphical presentations can be used to communicate information in relational data sets succinctly and effectively. However, novel graphical that represent many attributes relationships are often difficult understand completely until explained. Automatically generated must therefore either limited generating simple, conventionalized presentations, or risk incomprehensibility. A possible solution this problem would extend automatic presentation systems generate explanatory captions natural language, enable users the expressed graphic. This paper presents a system do so. It uses text planner determine content structure of based on: (1) representation its mapping it depicts, (2) framework for identifying perceptual complexity elements, (3) The output is further processed regarding issues such as ordering, aggregation, centering, referring expressions, lexical choice. We discuss architecture our strengths limitations. Our implementation currently 2-D charts maps, but, except information, domain independent. illustrate discussion with figures about housing sales Pittsburgh.