Traditional information extraction systems suffer from what Shasha Liao et al. term "tunnel vision": they analyze sentences or single documents in isolation. When a system processes a large corpus—such as a collection of news feeds, scientific literature, or historical memoirs—it inevitably encounters the same real-world event described multiple times across different sources. Without a mechanism to link these descriptions, the system treats every textual mention as a completely unique occurrence.
This fragmentation causes several critical failures across downstream applications:
- Data Redundancy and Overcounting: In automated database construction, such as the Gun Violence Database introduced by Ellie Pavlick et al., treating every news report as a separate incident distorts statistical analyses and skicking-point metrics by double-counting single events reported by multiple outlets.
- Incomplete Timelines: As noted by Ning Qiang et al., a lack of coreference resolution prevents temporal systems from realizing that synonymous phrases like "the blast" and "the explosion" point to the exact same chronological moment, resulting in fragmented and disjointed timelines.
- Information Silos: Valuable details are often distributed across different documents or languages. For instance, Mara Schilling-Wilhelmi et al. highlight how scientific data remains trapped because a material described in one paper cannot be automatically linked to experimental conditions detailed in another.
To solve this, the research direction moves toward Cross-Document Event Coreference Resolution, a paradigm designed to synthesize fragmented information into a unified, coherent representation.
Mentions vs. Instances
The core intuition of this field relies on a fundamental distinction formalized by Egoitz Laparra et al.: the difference between an event mention and an event instance.
- An event mention is the literal text snippet found within a document (e.g., "the company acquired the startup").
- An event instance is the canonical, real-world occurrence represented as a unique node in a knowledge graph or database.
By mapping multiple disparate mentions to a single instance, systems can aggregate complementary details—such as who, what, where, and when—from diverse sources. This allows for what Jun Zhou describes as Cross-Document Event Extraction (CDEE), shifting the focus from isolated sentence processing to a complete, unbiased "concept-level" view of complex real-world events.
How Identity is Established
How can an algorithm determine if two separate text snippets refer to the same real-world event, especially when the wording is entirely different or spans multiple languages? The literature offers several key mechanisms to build this intuition.
Co-Reporting and Record Linkage
Instead of relying purely on complex, ambiguous semantic boundaries, Joel Nothman et al. propose the principle of co-reporting. Under this framework, two mentions are considered coreferent if they both refer back to the same canonical archival story that first reported the incident. This transforms an abstract semantic problem into a functional search and retrieval task. Similarly, Timo Reuter et al. frame event identification as a record linkage problem, allowing systems to apply mature duplicate-detection techniques to cluster unstructured social media artifacts into coherent real-world events.
Centroid Arguments and Contextual Bridges
Often, reports share no common terminology. To bridge this gap, systems leverage shared anchors. Heng Ji et al. introduce the concept of "the centroid argument"—tracking highly salient individuals or entities across an entire corpus to automatically construct coherent event chains.
In a different domain, Jinyang Liu et al. demonstrate that system-side event logs can serve as a common contextual bridge. By correlating semantically diverse user support tickets (e.g., a "VM error" and a "Web App failure") to the same underlying system incident log, the framework resolves duplicates even when the textual descriptions share no common keywords.
Improving Extraction via Global Inference
Resolving coreference does not just clean up data after extraction; it actively improves the extraction process itself through global consistency constraints.
For example, Shasha Liao et al. implement a two-stage framework utilizing document-level cross-event relationships. Recognizing that an "Attack" event in one paragraph drastically increases the probability of a "Die" event in another allows the model to resolve ambiguities that single-sentence models miss. To scale this across longer distances, Xinya Du et al. introduce a Dynamic Global Memory framework, maintaining a document-level memory store to track event participants even when they are separated by many sentences from the event trigger.
Ultimately, these connections allow systems to move from isolated extraction tuples to what Rachel Wities et al. describe as an Open Knowledge Representation (OKR)—a unified graph that merges redundant or overlapping assertions across multiple texts into a single, comprehensive scenario.
Key Foundational Work
Cross-document Temporal and Spatial Person Tracking System Demonstration (2009) — contributed 2 concept elements to this direction.
Using Document Level Cross-Event Inference to Improve Event Extraction (2010) — contributed 1 concept elements to this direction.
Cross-document Event Extraction and Tracking: Task, Evaluation, Techniques and Challenges (2009) — contributed 1 concept elements to this direction.