Unsupervised Title Generation from Unpaired Data with N-Gram Discriminators

A title provides readers a succinct representation of the whole document and consequently helps them get the gist without going through the details. Although the research of title generation has been investigated for decades, it still leaves much space to improve due to its requirements of complex tasks such as language understanding and text synthesis. To perform unsupervised learning of titles, we employ a generative adversarial network (GAN) without paired training data. A sequence-to-sequence model with an auto-encoder architecture is adapted to model the generator. The readability gap between human-written and machine-generated text still needs to reduce. As n-gram model has been successfully used as a language model for solving the language tasks, we present n-gram discriminators in the GAN model to increase the readability of machine-generated titles. The proposed model is tested on the English Gigaword dataset and the experimental results demonstrate that our approach obtains encouraging results.