Abstract: Ling Luo, Xiang Ao, Yan Song, Feiyang Pan, Min Yang, Qing He. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).
2019.
Introduction
Extractive summarization—the process of identifying and concatenating the most representative sentences from a document—is a cornerstone of natural language processing. While generally more reliable and grammatically consistent than abstractive methods, current state-of-the-art models struggle with two primary issues: "Lead Bias" and sequential limitations. Lead bias refers to the tendency of models to over-rely on the first few sentences of a text (the "Lead-3" phenomenon), often missing critical information buried deeper in long-form content. Furthermore, standard sequence-labeling approaches often treat sentences in isolation or strict linear order, failing to capture the global and paragraph-level context essential for determining true importance.
To address these challenges, the authors introduce HER (Human-bEing-Reading), a novel framework that mimics the cognitive process humans use to digest long documents: a "rough reading" phase to capture the gist and a "careful reading" phase to select key details.
Methodology: The HER Framework
The HER model decomposes the summarization task into two distinct stages, as illustrated in the architecture below:
1. Rough Reading (Encoding)
In this stage, the model builds a "general cognition" of the document using two feature types:
- Global Features: A hierarchical BiLSTM encodes sentences and the entire document into a global vector.
- Local Features: A Convolutional Neural Network (CNN) is applied over sentence embeddings to capture the "gist" of different paragraphs. This allows the model to understand local context rather than just individual sentence semantics.
2. Careful Reading (Selection)
The selection process is formulated as a Contextual Bandit problem. An RL agent observes the document context and selects sentences iteratively based on three principles: Salience (importance), Coverage (matching paragraph gists), and Redundancy (avoiding repetition).
Unlike models that extract a fixed number of sentences, HER employs an Adaptive Termination mechanism. Using a Bernoulli-based policy, it stops extraction when the affinity of remaining sentences drops below a threshold relative to the already selected content. The model is trained using the REINFORCE algorithm with a self-critical baseline to optimize directly for ROUGE scores.
Experimental Results
HER was evaluated on the CNN, DailyMail, and combined CNN/DailyMail datasets (approximately 300,000 pairs).
Performance and Distribution
HER achieved a state-of-the-art ROUGE-1 score of 42.3 on the combined dataset, outperforming previous reinforcement learning models like BANDITSUM (41.5). Crucially, HER successfully mitigated lead bias. As shown in the distribution analysis below, HER selects sentences from across the entire document length, whereas baselines remain heavily clustered at the beginning.
Human Evaluation
In blind tests, human volunteers ranked HER as the top-performing system in terms of overall quality, coverage, and non-redundancy, confirming that the human-inspired reading process translates to higher-quality summaries.
Discussion and Future Work
While highly effective, HER has limitations. The CNN-based local feature extractor requires a minimum document length (at least 3 sentences), making it less suitable for very short snippets. Additionally, the two-stage RL training process is more computationally intensive than simple cross-entropy labeling.
Future research may adapt this multi-stage reading process for abstractive summarization or Machine Reading Comprehension (MRC) tasks, potentially bringing human-like cognitive strategies to a broader range of generative and analytical NLP problems.
Summary updated