Summarization with deep transfer learning

Abstract

There have been many successes in recent neural network-based approaches for summarization. Despite the impressive results they achieved, these methods have their limitations. Previous neural methods for extractive summarization focus on improving the saliency of the extracted sentences. However, they fail to consider coherence, and hence sometimes produce unreadable summaries. We propose a coherence-reinforced extractive summarization model, which has two parts: an extractive summarization model and a coherence model. The summarization model learns to maximize coherence and saliency simultaneously, by transferring coherence awareness from the coherence model to the summarization model via reinforcement learning. The experimental results show that the proposed model outperforms previous works in both ROUGE scores and human evaluation. For abstractive summarization, most neural approaches require a considerable amount of training data. However, training data is insufficient in domains such as Femail, ScienceTech, and Health, and abstractive summarization models perform poorly on these domains. To alleviate this problem, we propose to adopt transfer learning methods to abstractive summarization, so that the model could exploit summarization knowledge learned from large domains. The experimental results demonstrate that the introduction of transfer learning significantly improves the performance of abstractive summarization models on low-resource domains.

Date of Award	2018
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology