An Evolutionary Approach to Referring Expression Generation and Aggregation

The work presented here is intended as an evolutionary task-specific module for referring expression generation and aggregation to be enclosed in a generic flexible architecture. Appearances of concepts are considered as genes, each one encoding the type of reference used. Three genetic operators are used: classic crossover and mutation, plus a specific operator dealing with aggregation. Fitness functions are defined to achieve elementary coherence and stylistic validity. Experiments are described and discussed.