When a complex news story or crisis unfolds, information does not arrive all at once. Instead, it emerges as a continuous stream of updates, corrections, and shifting contexts. If you want to understand a long-running event, traditional multi-document summarization algorithms fall short.
The reason lies in how these traditional models are built: they treat a collection of documents as a static pool, optimizing for "centrality"—the information that is most frequently repeated across the texts. However, in an evolving narrative, what was central yesterday becomes redundant background information today. As Kathleen McKeown et al. observe in their development of the Newsblaster system, users tracking ongoing stories are frequently bogged down by repetitive context (Columbia's Newsblaster: New Features and Future Directions). To solve this, the field of update summarization shifts the goal from finding what is most common to finding what is incrementally new.
The Relevance-Novelty Tradeoff
The core challenge of update summarization is balancing two competing forces: relevance and novelty. A good update summary must contain information that is highly relevant to the core topic, but it must strictly exclude anything the reader has already seen in previous articles.
This is harder than it sounds. You cannot simply apply a standard redundancy penalty to words that appeared in past documents. As Melissa Egan et al. demonstrate, basic redundancy metrics fail to capture the nuance of how a story develops over time; they often accidentally penalize the actual updates because new facts are naturally intertwined with old vocabulary (Multiple Alternative Sentence Compressions and Word-Pair Antonymy for Automatic Text Summarization and Recognizing Textual Entailment).
To solve this mathematically, systems must model a user's prior knowledge state. Annie Priyadarshini Louis introduced a Bayesian framework that explicitly incorporates historical background knowledge as a prior distribution (A Bayesian Method to Incorporate Background Knowledge during Automatic Text Summarization). This allows the system to calculate the marginal utility of a sentence—estimating how much new information it provides given what is already known.
Standard Summarization: Maximize Relevance(Current Documents)
Update Summarization: Maximize Relevance(Current Documents) - Redundancy(Prior Knowledge)
Other approaches handle this tension structurally. Xiaojun Wan et al. proposed a co-ranking framework that applies a mutual exclusivity constraint between an "update score" and a "consistency score" (Update Summarization Based on Co-Ranking with Constraints). This ensures that sentences containing vital new developments are pushed to the top, while sentences that merely reiterate established facts are suppressed. Interestingly, when evaluating these architectures, Shafiq Joty et al. discovered that while complex supervised models excel at identifying what is broadly important, simpler text-overlap metrics are often highly effective at the specific task of filtering out historical background knowledge (UofL at TAC 2008 Update Summarization and Question Answering).
Anchoring the Shifting "Now"
To determine what constitutes an "update," a system must accurately sequence events in time. This introduces a major linguistic hurdle: temporal resolution.
News articles rarely use absolute calendar dates. Instead, they rely on relative expressions like "yesterday," "last week," or "the following month." If an algorithm assumes a static reference point—such as the document's publication date—it will quickly lose track of the timeline when a narrative flashes backward or forward.
Gabor Angeli et al. point out that because the narrative focus constantly shifts within a text, a system must dynamically update its internal anchor of the "now" point to resolve dates accurately (Language-Independent Discriminative Parsing of Temporal Expressions).
The mechanics of this require tracking a dynamic temporal focus. As pioneered by Inderjeet Mani et al., systems can combine discourse context with verb tenses to create a moving reference window (Robust Temporal Processing of News). For example, consider the sequence:
"The storm hit on Tuesday. Officials expected power to return the next day."
By tracking the temporal focus, the system anchors "the next day" to Wednesday, rather than the date the article was published. Kenton Lee et al. formalized this using context-dependent semantic parsing, demonstrating that modeling these shifting reference points drastically reduces errors when reconstructing event sequences from narrative text (Context-dependent Semantic Parsing for Time Expressions).
Mapping the Lifecycle of an Event
Once a system can filter out old information and accurately anchor time, it can begin to map the broader lifecycle of an event. Information streams do not just accumulate linearly; sub-topics emerge, merge, and fade away.
To capture this fluid progression, researchers have moved away from static topic modeling toward evolutionary architectures. Jiwei Li et al. developed the Evolutionary Hierarchical Dirichlet Process (EHDP), which links non-parametric topic models across time-dependent steps (Evolutionary Hierarchical Dirichlet Process for Timeline Summarization). This mechanism allows an AI to mathematically detect the exact moment a new sub-event is born or when an old thread dies.
[Time T1: Sub-topic A] ---\
===> [Time T2: Merged Topic C] ---> [Time T3: Topic Dies]
[Time T1: Sub-topic B] ---/
For massive document collections where broad topic models lack the granularity to track specific incidents, researchers use document "threading." Jennifer Gillenwater et al. introduced a framework using Structured Determinantal Point Processes (SDPPs) to extract distinct, parallel narrative chains that are individually coherent but collectively diverse (Discovering Diverse and Salient Threads in Document Collections).
This continuous, granular tracking is particularly vital in high-stakes scenarios like natural disasters or local crises. As Hossein Rajaby Faghihi et al. highlight in their work on crisis benchmarks, capturing the real-time lifecycle of an event prevents critical updates from being swallowed by the noise of the surrounding media storm (CrisisLTLSum: A Benchmark for Local Crisis Event Timeline Extraction and Summarization). By treating summarization as a dynamic, historical process rather than a static snapshot, update summarization provides a clear window into how our world changes sentence by sentence.
Key Foundational Work
Robust Temporal Processing of News (2000) — contributed 2 concept elements to this direction.
Summarizing Time-Varying Data (1997) — contributed 2 concept elements to this direction.