Smart Initialization Yields Better Convergence Properties in Deep Abstractive Summarization

Abstractive text summarization has been proposed as an alternative to the inherently limited extractive methods, but extant work is plagued with high training times. In this work, we introduce a set of extensions, including novel initialization techniques, that allow contemporary models to achieve 10x faster training time and comparable results. Our work also provides substantial evidence against the accepted evaluation metric for abstractive summarization, and establishes a speed benchmark for further research.ive text summarization has been proposed as an alternative to the inherently limited extractive methods, but extant work is plagued with high training times. In this work, we introduce a set of extensions, including novel initialization techniques, that allow contemporary models to achieve 10x faster training time and comparable results. Our work also provides substantial evidence against the accepted evaluation metric for abstractive summarization, and establishes a speed benchmark for further research.