A Dynamic Programming Approach to Document Length Constraints

Natural language generation (NLG) applications must occasionally deliver rhetorically coherent output under length constraints. For example, certain types of documents must fit on a single webpage, on a cell phone screen, or into a fixed number of printed pages. To date, applications have achieved this goal by structuring their content as a rhetorical tree and using a greedy algorithm to pick the discourse elements to include in the final document. Greedy algorithms are known to pick sub-optimal solutions. This paper presents an alternate approach based on dynamic programming.