PSO Optimal Parameters and Fitness Functions in an NLP Task

The present article investigates the optimal setup of a particle swarm algorithm for a specific natural language processing task. This task consists of identifying syntactic phrases in unconstrained natural language texts, by using attractive and repulsive forces between neighbouring words. Here, a number of avenues for improving the swarm effectiveness are investigated, clustered around two main research directions. The first direction involves the optimal choice of main swarm parameters, namely the inertia, cognitive and social parameters. The research question is if the values of these parameters may be selected to give a substantial improvement in the optimization process, and whether the best values recommended for other tasks are effective in the chosen task. The second direction involves the choice of fitness function which determines the fitness of each solution, and guides the optimization process. Lesser used fitness functions are applied, to determine whether they can generate an improved phrasing solution. Experiments show these two research directions to lead to a substantial improvement in the phrasing accuracy.