Improving Abstractive Summarization by Transfer Learning with Adaptive Document Selection

Abstract

ive document summarization based on neural networks is a promising approach to generate a flexible summary but requires a large amount of training data.While transfer learning can address this issue, there is a potential concern about the negative transfer effect that deteriorates the performance when we use training documents irrelevant to the target domain, which has not been explicitly explored in document summarization tasks.In this paper, we propose a method that selects training documents from the source domain that are expected to be useful for the target summarization.The proposed method is based on the similarity of word distributions between each source document and a set of target documents.We further propose an adaptive approach that builds a custom-made summarization model for each test document by selecting source documents similar to the test document.In the experiment, we confirmed that the negative transfer actually happens also in the document summarization tasks.Additionally, we show that the proposed method effectively avoids the negative transfer issue and improves summarization performance.