Word difficulty prediction—based on GA-BP neural network and TOPSIS-GRA method

Abstract

The game WORDLE is very popular around the world, so our team tried to build a suitable model to analyze the difficulty of the words in the game. Firstly, this paper preprocesses the data to ensure the accuracy and integrity of the data. After extracting five representative characteristics, such as "letter frequency score", "word distance" and "syllable harmony", GA-BP neural network was selected to predict the percentage distribution of word attempts. The results show that compared with the traditional BP neural network, the GA-BP model with learning rate of 0.3, crossover rate of 0.7 and variation rate of 0.01 maintains a higher prediction accuracy, and the MSE value is significantly reduced. In addition, this paper combines GRA model and TOPSIS method, and uses entropy weight method to calculate the objective weight of indicators to comprehensively evaluate the difficulty of words. The example analysis shows that the improved model is more distinguishable and overcomes the defect of TOPSIS method that Euclidean distance can’t distinguish ranking. At the same time, the ranking results of the two are close, which indicates the reliability and rationality of the difficulty prediction by using the model.