Content planner construction via evolutionary algorithms and a corpus-based fitness function
Kathleen R. McKeown, Pablo Ariel Duboue · 2002
In this paper, we present a novel technique to learn a tree-like structure for a content planner from an aligned corpus of semantic inputs and corresponding, human-produced, outputs. We apply a stochastic search mechanism with a two-level fitness function. As a first stage, we use high level order constraints to quickly discard unpromising planners. As a second stage, alignments between regenerated text and human output are employed. We evaluate our approach by using the existing symbolic planner in our system as a gold standard, obtaining a 66 % improvement over a random baseline in just 20 generations of genetic search.