Evolutionary and Case-Based Approaches to REG: NIL-UCM-EvoTAP, NIL-UCM-ValuesCBR and NIL-UCM-EvoCBR

We propose the use of evolutionary algorithms (EAs) (Holland, 1992) to deal with the attribute selection task of referring expression generation. Evolutionary algorithms operate over a population of individuals (possible solutions for a problem) that evolve according to selection rules and genetic operators. The fitness function is a metric that evaluates each of the possible solutions, ensuring that the average adaptation of the population increases each generation. Repeating this process hundreds or thousands of times leads to very good solutions for the problem. We encode as a fitness function the specific constraints required for the reference to be acceptable. The crossover and mutation genetic operators ensure a reasonable variation between the different options much as a human-generated text would. Each individual is represented by a set of genes that are the list of possible attributes in the reference. Each gene has an associated value of 0 (if the attribute is not included in the reference), or 1 (if the attribute is included in the reference). The initial population should have a low number of genes set to 1, because references tend to be short and the use of all the possible attributes should be avoided. For the crossover operator, two individuals are selected randomly and crossed by a random point of their structure. For the mutation operator, some of the genes are chosen randomly to be mutated from 1 to 0, or vice versa. The fitness function must find a balance between the univocal identification of a referent, and a natural use of attributes. The formula used as fitness function is defined in Equation 1: fitindi = fatti∗weightatt+ident∗weightid (1)