Research On Modified Focal Loss Function Of Intent Detection

The dialogue system is one of the most important research directions in natural language processing(NLP).It also is the foundation in the field of smart phones and smart home.Intent detection also plays an important role in pipeline dialogue system.At present,although the field of intent detection has made great progress,there are still some problems and challenges.For example,there are data imbalance in intent detection field.If the problem aboved can be solved,the performance of dialogue system can be further improved.In order to force on the problems above,the research contents include as follow:1.This research introduces the concept of decline remaining which reflect difference between cross entropy loss function and the modified version based on cross-entropy.The value of focal loss is always lower than the value of cross-entropy.When the probability of the model predicts correctly higher,the decline remaining is lower.It reflects the limitation of the focal loss in dealing with imbalance data.2.This research preprocesses the Modified Focal Loss and introduces the decline threshold in the loss function.This hyper-parameter is used to control the loss value is decline or extent at a certain probability compared to cross-entropy.Based on the concept of decline threshold,the larger of the probability,the lower of the decline remaining in Modified Focal Loss.3.Based on the cross-entropy loss function,Focal Loss loss function and the Modified Focal Loss loss function,the experiments show that the proposed new loss function is effective and feasible for the classification effect of unbalanced data.