Gradient Optimization of Beam Search Hyperparameters

Conditional text generation models based on neural networks are currently widely used in applied natural language processing. Beam search is a broadly used method for generating a prediction with the highest estimated probability at test time. However, it was shown to produce suboptimal results without introducing refinements to the score function, which entails the need to use hyperparameters balancing the influence of each factor on the hypothesis score. Currently, the problem of finding the set of hyperparameters yielding the highest value of the target metric during validation is mainly solved by grid search. This method has exponential complexity with respect to the number of unknown variables, which raises the question of the existence of a more efficient approach. In this research, we propose a gradient method for finding the optimal values of beam search hyperparameters by incorporating them in the loss function which is minimized during training.