This code is meant as a starting point for implementing NLG solutions using genetic algorithms.

To run the sample code, you need the below requirements: pip install numpy deap mosestokenizer

• main.py: The main script to execute the genetic algorithms. The input is handled here along with any parsing/tokenizations. Parameters to the genetic algorithm (e.g. population size, number of generations) are defined here.
• genetic_algorithm.py: The implementation of the evolutionary algorithm.
• creative_code.py: Contains the mutation, crossover and evaluation functions
  • weights: A tuple containing the weights (between -1 to +1) of each optimization dimension. A positive weight indicates maximizing the dimension and negative is for minimizing it.
  • mutate: This function is in charge of mutating a single individual (e.g. changing one token with a one). The function should return a tuple, so the comma after the individual is needed.
  • crossover: This function takes two individuals. It picks a point at random and swaps tokens (that are before the selected point) of both individuals.
  • evaluate_ind: A function that evaluates an individual. The function receives the current individual and the original individual that was set in main.py. The function must return the same number of scores as the number of dimensions set by weights.

The code could be extended to use additional information of the input (such as dependency parse-trees, themes and such). For that, the genetic_algorithm.py and main.py would need to be modified accordingly.

Methods that have employed evolutionary algorithms to generate creative language (in English and Finnish):

• Khalid Alnajjar and Hannu Hannu Toivonen (2020). Computational Generation of Slogans. Natural Language Engineering.
• Mika Hämäläinen and Khalid Alnajjar (2019). Let's FACE it. Finnish Poetry Generation with Aesthetics and Framing. In the Proceedings of The 12th International Conference on Natural Language Generation. pages 290-300.
  • GitHub: FinMeter