[Paper Review] Fast Abstractive Summarization with Reinforce-Selected Sentence Rewriting

Yen-Chun Chen, Mohit Bansal|arXiv (Cornell University)|May 28, 2018

Topic ModelingComputer Science40 references94 citations

TL;DR

The paper presents a hybrid extractive-abstractive summarization model that first selects salient sentences with a reinforcement-learning extractor and then rewrites them abstractively, achieving state-of-the-art ROUGE/METEOR on CNN/Daily Mail and faster inference via parallel sentence decoding.

ABSTRACT

Inspired by how humans summarize long documents, we propose an accurate and fast summarization model that first selects salient sentences and then rewrites them abstractively (i.e., compresses and paraphrases) to generate a concise overall summary. We use a novel sentence-level policy gradient method to bridge the non-differentiable computation between these two neural networks in a hierarchical way, while maintaining language fluency. Empirically, we achieve the new state-of-the-art on all metrics (including human evaluation) on the CNN/Daily Mail dataset, as well as significantly higher abstractiveness scores. Moreover, by first operating at the sentence-level and then the word-level, we enable parallel decoding of our neural generative model that results in substantially faster (10-20x) inference speed as well as 4x faster training convergence than previous long-paragraph encoder-decoder models. We also demonstrate the generalization of our model on the test-only DUC-2002 dataset, where we achieve higher scores than a state-of-the-art model.

Motivation & Objective

• Motivate fast, accurate long-document summarization by combining extractive saliency with abstractive rewriting.
• Develop a sentence-level RL bridge to connect extractor and abstractor without sacrificing fluency.
• Achieve state-of-the-art performance on CNN/Daily Mail and generalize to DUC-2002.
• Enable parallel decoding to dramatically speed up training and inference.

Proposed method

• Two-module architecture: an extractor that selects salient sentences and an abstractor that rewrites them.
• Hierarchical sentence representations via temporal convolution and bidirectional LSTM for context.
• Sentence selection with a Pointer Network and a two-hop attention mechanism, trained via policy gradient (A2C) with ROUGE-based rewards.
• Abstractor as encoder-aligner-decoder with copy mechanism to generate concise, fluent rewrites.
• ML pre-training for extractor (proxy saliency labels) and abstractor (sentence-level pairs); RL (A2C) to fine-tune extractor end-to-end while keeping abstractor language model fixed.
• Optional reranking to reduce across-sentence redundancy using tri-gram avoidance and diverse decoding.

Experimental results

Research questions

• RQ1 Can a sentence-level RL extractor effectively identify salient sentences for abstractive rewriting?
• RQ2 Does a coarse-to-fine, extract-then-rewrite approach improve quality, speed, and fluency compared to end-to-end long-document encoders?
• RQ3 What are the gains in abstractive quality and novelty when combining RL-guided extraction with a copying-enabled abstractor?
• RQ4 How well does the model generalize to out-of-domain datasets like DUC-2002?

Key findings

• Achieves new state-of-the-art ROUGE and METEOR on CNN/Daily Mail for both extractive and abstractive settings.
• RL-guided extractor with parallel sentence-level decoding yields 10-20x faster inference and up to 4x faster training convergence than prior long-encoder models.
• Abstractive model generates substantially more novel N-grams (3x more) than strong baselines, indicating true abstractive rewriting.
• Outperforms a state-of-the-art model on the test-only DUC-2002 dataset, showing good generalization.
• Human evaluation shows improved relevance and readability versus a strong prior model.
• Optional reranking further reduces across-sentence redundancy with measurable gains in ROUGE.