Reproduce “A Cascade Approach to Neural Abstractive Summarization with Content Selection and Fusion” on OpenResearch

About this paper

Logan Lebanoff, Franck Dernoncourt, Doo Soon Kim, Walter Chang, Fei Liu

Researchers at the University of Central Florida and Adobe Research developed a cascade architecture for neural abstractive summarization that explicitly separates content selection and text generation. This modular approach, which employs a coarse-to-fine content selection strategy and a neural fusion model, achieved ROUGE scores competitive with end-to-end systems on the CNN/Daily Mail dataset and highlighted content selection as the primary area for future performance gains.

Key insights

• A modular cascade approach offers a viable and competitive alternative to end-to-end neural abstractive summarization, particularly for adapting to domains with limited data.
• The accuracy of content selection, at both sentence and fine-grained segment levels, represents the most significant bottleneck and holds the largest potential for improving abstractive summary quality.
• Explicitly guiding the neural generator with fine-grained highlight tags, integrated as embeddings, effectively steers surface realization, enhancing faithfulness and coherence in generated summaries.

Summary by alphaXiv.