Abstractive Summarization Model with Adaptive Sparsemax

The Abstractive summarization models mostly rely on Sequence-to-Sequence architectures, in which the softmax function is widely used to transform the model output to simplex. However, softmax's output probability distribution often has the long-tail effect especially when the vocabulary size is large. Many unrelated tokens occupy too many probabilities so they will reduce the training efficiency and effect. More recently, some work has begun to design mapping functions to gain sparse output probabilities to ignore these irrelevant tokens. In this paper, we propose Adaptive Sparsemax which can self-adaptively control the sparsity of the model's output. Our method combines sparsemax and temperature mechanism, and the temperature value can be learned by the neural network. One of the advantages of our method is that it doesn't need any hyperparameter. The experimental result on CNN-Daily Mail and LCSTS dataset shows that our method has better performance on the abstractive summarization task than baseline models.