When an Information Extraction (IE) system processes a single sentence, it faces an uphill battle. Natural language is inherently messy, ambiguous, and full of complex syntactic structures. To extract a fact reliably from an isolated sentence, a local extractor must achieve near-perfect recall and precision simultaneously—a daunting task for automated systems.
However, when scaling up to massive text corpora like the World Wide Web, news archives, or biomedical literature, a powerful property emerges: information redundancy. Important facts are rarely stated just once. Instead, they are repeated across different documents, written by different authors, and expressed through diverse linguistic structures. This global repetition shifts the problem from deep linguistic analysis of isolated sentences to statistical aggregation across an entire corpus.
Flipping the Tradeoff: High Precision, Low Recall
The most immediate benefit of information redundancy is that it allows system designers to intentionally build high-precision, low-recall extraction mechanisms. In a traditional small-scale setting, a low-recall system is useless because it misses too much information. But in a massive corpus, the probability of capturing a fact increases with the number of times it is mentioned.
As shown by Michael Fleischman et al. in Offline Strategies for Online Question Answering: Answering Questions Before They Are Asked, processing massive text archives enables systems to prioritize strict, high-precision rules. Even if a specific pattern misses 90% of a fact's occurrences, the frequent recurrence of core facts across diverse sources ensures that the target information is eventually captured.
This strategy is particularly vital in specialized domains like biomedicine, where data quality is paramount. For instance, K. Bretonnel Cohen et al. demonstrated in High-precision biological event extraction with a concept recognizer that relying on strict, ontologically typed rules yields top-tier precision because core biological facts are frequently repeated across different publications.
Similarly, Junkyu Lee et al. in On the efficacy of per-relation basis performance evaluation for PPI extraction and a high-precision rule-based approach and Jari Björne et al. in Scaling up Biomedical Event Extraction to the Entire PubMed leveraged this exact property to scale extraction to millions of abstracts. By focusing only on highly confident, unambiguous extractions, these systems maintain impeccable data quality, trusting that the sheer volume of the corpus will compensate for individual missed instances.
Traditional Local IE:
[Single Sentence] ───► [Complex Parser] ───► High Risk of Error (Low Precision)
Redundancy-Driven IE:
[Sentence Mention A] ──┐
[Sentence Mention B] ──┼─► [Strict, High-Precision Rules] ──► Confirmed Fact (High Precision)
[Sentence Mention C] ──┘
Error Correction and the Wisdom of the Crowd
Redundancy does more than just recover missing facts; it acts as an automated error-correction mechanism. If a noisy sentence-level parser extracts an incorrect fact, that error is likely to be a sporadic, isolated event. Genuine facts, conversely, will cluster together as multiple distinct linguistic expressions pointing to the same underlying relationship.
Roman Yangarber et al. explored this in Redundancy-based Correction of Automatically Extracted Facts, showing that aggregating knowledge from a large collection of extracted facts allows a system to identify and rectify inaccuracies that look plausible in isolation. Instead of relying on local single-document analysis, the global distribution of facts provides the necessary context for validation.
This cross-document evidence aggregation is highly effective for web-scale mining. Benjamin Rosenfeld et al. in Using Corpus Statistics on Entities to Improve Semi-supervised Relation Extraction from the Web and Vivi Nastase et al. in Semantic Relations Between Nominals detailed how models can cross-reference multiple linguistic expressions to filter out noise. By treating the web as a self-correcting corpus, systems can favor facts that appear consistently across different sources. In the thesis Web relation extraction with distant supervision, Isabelle Augenstein formalized this by aggregating multiple extraction predictions using knowledge base statistics, proving that post-hoc integration compensates for individual extraction errors.
Triangulation via Mutual Reinforcement
How do we harness this redundancy without human supervision? One elegant solution is to model the relationship between extraction patterns and extracted facts as a bipartite graph.
In Unsupervised Information Extraction Approach Using Graph Mutual Reinforcement, Hany Hassan et al. proposed an unsupervised framework that treats extraction templates as "authorities" and extracted data points as "hubs" (similar to web search ranking algorithms).
If a fact is extracted by many different reliable patterns, the fact is likely true. Conversely, if a pattern successfully extracts many known true facts, the pattern is likely reliable. This mutual reinforcement exploits the inherent repetition in large-scale text corpora to triangulate validity without requiring human-labeled seed examples.
Extraction Patterns Extracted Facts
┌─────────────────────┐ ┌───────────────────┐
│ "X is author of Y" │ ───┐ ┌───►│ (G. Orwell, 1984)│
└─────────────────────┘ │ │ └───────────────────┘
├───┤
┌─────────────────────┐ │ │ ┌───────────────────┐
│ "Y, written by X" │ ───┘ └───►│ (H. Melville, MD)│
└─────────────────────┘ └───────────────────┘
(High Overlap = High Trust) (Repeated = Validated)
More recently, Shuai Ma et al. introduced the eFIRE framework in Easy First Relation Extraction with Information Redundancy, which uses an "easy-first" strategy. By first extracting facts from sentences where the relationship is glaringly obvious and unambiguous, the model establishes a foundation of trusted facts. It then leverages corpus-level redundancy to resolve more difficult, ambiguous sentences later in the pipeline.
Expanding Beyond Standard Corpora
The intuition of redundancy-driven extraction extends beyond web pages and scientific papers into structured technical domains and open-domain question answering.
For example, technical specifications and network protocols often repeat data field properties across multiple sections. Samuel Jero et al. demonstrated in Leveraging Textual Specifications for Grammar-Based Fuzzing of Network Protocols that aggregating information from these various redundant mentions ensures that properties missed in one context are captured in another, creating highly robust protocol grammars automatically.
In open-domain question answering, traditional systems often struggle to consolidate evidence when a candidate answer appears across multiple distinct passages. As Bernhard Kratzwald et al. noted in RankQA: Neural Question Answering with Answer Re-Ranking, without a mechanism to aggregate these occurrences, models miss a critical signal for verification. By introducing a re-ranking stage that fuses information retrieval features with machine comprehension, RankQA explicitly counts and leverages these redundant answer occurrences to defend against the noise inherent in large-scale document collections.
Key Foundational Work
Offline Strategies for Online Question Answering: Answering Questions Before They Are Asked (2003) — contributed 2 concept elements to this direction.
Redundancy-based Correction of Automatically Extracted Facts (2005) — contributed 2 concept elements to this direction.