Main Point Generator: Summarizing with a Focus

Text summarization is attracting more and more attention while deep neural network has had many successful application in NLP. One problem of such models is its inability to focus on the essentials of documents, thus generating summaries that may not be important, especially during multi-sentence summarization. In this paper, we propose Main Pointer Generator (MPG) to address the problem, where at each decoder step the whole document is taken into consideration when calculating the probability of next generated token. We experiment with CNN/Daily news corpus and results show that summaries our MPG generated follow the main theme while outperforming the original pointer generator network by about 0.5 ROUGE point.