摘要作為人們快速了解資訊的手段，一直以來都是自然語言處理研究的主要方向之一。現今的摘要模型主要都是依靠深度學習模型，讓模型自己決定文章的重點以及摘要生成的內容，因此人為可控制的因素較小。而本論文認為在某些摘要的應用場景中，摘要的重點不應該只依靠模型本身決定，而需要一些其他的資訊來輔助模型產生更貼近文章重點的摘要。最終，我們在現有摘要模型的輸入上做一些改動，使其能夠產生相對應內容的摘要。除此之外，我們也針對資訊擷取模型進行遷移式學習，使其能更適合應用於我們的使用場景。;Abstract is the main method that help people quickly understand the information of the article, and it is also a main research topic of Natural Language Processing. Modern abstractive summarization model mainly relies on deep learning methods, and need model itself to determine the key point of the article and the content of the abstract, there few human control factors in it. In this paper, we believe that in some scenarios of summarization, the content of the abstract should not only rely on model itself, we need to give more additional information to help model generate topic related abstract. Finally, we modify the input of the model to allow it generate the abstract with corresponding content. Additionally, we apply transfer learning on existing information extraction model to help it more suitable in our scenario.
