In extractive summarization, the goal is to select a subset of sentences from one or more documents that best represents the source material. A naive approach might simply rank sentences by their relevance to a query or their centrality to a document and pick the top $N$. However, this often leads to an "echo chamber" effect: in a cluster of news articles about the same event, the most "relevant" sentences are often nearly identical, leading to a summary that repeats the same fact multiple times while missing other key details.
To solve this, researchers use Redundancy-Aware Selection, a mechanism that treats the value of a sentence not as a fixed score, but as a dynamic property that changes based on what has already been selected.
The Balancing Act: Relevance vs. Novelty
The core intuition behind this research direction is a trade-off between two competing objectives: relevance (how well a sentence covers the target topic) and novelty (how much new information it adds to the existing summary).
As You et al. in A Study on Position Information in Document Summarization and Lebanoff et al. in Scoring Sentence Singletons and Pairs for Abstractive Summarization explain, this is frequently formalized as Maximal Marginal Relevance (MMR). MMR iteratively selects the candidate that maximizes the difference between its relevance score and its maximum similarity to the sentences already in the summary. By doing so, the system ensures that each addition provides diverse coverage of the source document rather than repeating content overlap.
The Greedy Selection Loop
The most common implementation of this idea is a greedy iterative algorithm. Instead of calculating an optimal set of sentences all at once—which is computationally expensive—the system builds the summary one step at a time:
- Rank all candidate sentences based on their initial importance or relevance.
- Select the highest-ranked sentence and add it to the summary.
- Penalize the remaining candidates based on their similarity to the selected sentence.
- Repeat until the summary reaches a desired length.
Wan et al. demonstrate this in Manifold-Ranking Based Topic-Focused Multi-Document Summarization, noting that purely ranking-based selection in graph models often fails because similar sentences receive similar scores. By applying a greedy diversity penalty, the scores of unselected sentences are dynamically updated, ensuring the final output is both salient and diverse. Chandu et al. apply a similar iterative process in Tackling Biomedical Text Summarization: OAQA at BioASQ 5B to bridge the gap between technical questions and concise "ideal answers."
Mechanisms for Penalizing Redundancy
Researchers have proposed various ways to define and calculate the "penalty" for redundancy:
- Word-Level Penalties: Nenkova in Entity-driven Rewrite for Multi-document Summarization describes a method where the weights of words are "zeroed out" once they are included in a summary. Similarly, He et al. in Exploring Differential Topic Models for Comparative Summarization of Scientific Papers penalize the probability of words already selected to ensure the summary covers a broader range of unique concepts.
- Similarity Thresholds: Some systems use a hard filter. Garg et al. in Clusterrank: a graph based method for meeting summarization and Yasunaga et al. in Graph-based Neural Multi-Document Summarization use similarity thresholds, where a candidate is only included if its overlap with the existing summary remains below a certain limit.
- Semantic and Intent Coverage: In more complex tasks, like answering essay questions, Botros et al. in CMUQA in QALab-3: Essay Question Answering for University Entrance Exams use diversification algorithms to ensure proportionality to different "intents," ensuring every required topic is addressed without repetition.
Applications Across Domains
This redundancy-aware framework is essential whenever information is dense or multi-sourced. In scientific summarization, Zerva et al. use it in Cited text span identification for scientific summarisation using pre-trained encoders to select diverse cited spans. In social media, Xu et al. in Topic-Guided Self-Introduction Generation for Social Media Users use redundancy-aware selection to distill a user's noisy tweet history into a representative bio. Even in high-velocity data streams, Chen et al. in Region-Based Message Exploration over Spatio-Temporal Data Streams show that pruning redundant candidates allows for high-quality summaries with significantly lower computational overhead.
Key Foundational Work
Multi-Answer-Focused Multi-Document Summarization Using a Question-Answering Engine (2004) — contributed 2 concept elements to this direction.
Corpus and Evaluation Measures for Multiple Document Summarization with Multiple Sources (2004) — contributed 1 concept elements to this direction.
QUALIFIER In TREC-12 QA Main Task (2003) — contributed 1 concept elements to this direction.