Traditional information retrieval systems treat text as a sequence of raw tokens. If a user searches for a specific concept, a keyword-based engine looks for exact character matches, while a standard dense retriever projects the text into a continuous vector space based on surface-level context. Both approaches suffer from fundamental limitations: they are vulnerable to vocabulary mismatch, struggle with homonyms and synonyms, and fail to leverage the structured relationships inherent in real-world knowledge.
Entity-centric information retrieval shifts the fundamental unit of search from text tokens to unique, grounded entities (such as specific people, places, organizations, or biological concepts). This paradigm recognizes that human intent and factual knowledge are organized around distinct objects rather than arbitrary strings of words. As Andrew Tomkins observes in The Future of Search: An Online Content Perspective, web content and user queries are increasingly directed toward structured entities rather than static, unstructured pages. Similarly, log analyses by Ravi Kumar et al. in A Characterization of Online Search Behavior and Katja Hofmann et al. in A Semantic Perspective on Query Log Analysis confirm that the vast majority of search traffic in both general and specialized domains targets named entities.
By indexing, expanding, and ranking documents based on the explicit identities of the entities they contain, this approach bridges the gap between unstructured text and structured knowledge graphs.
Grounding the Index in Identity
The core mechanism of entity-centric retrieval is the replacement of raw text tokens with unique entity identifiers within the search index.
In traditional text retrieval, synonyms (e.g., "cancer" vs. "neoplasm") create terminological gaps, while homonyms (e.g., "Apple" the company vs. "apple" the fruit) introduce noise. By mapping text mentions to a canonical knowledge base via Named Entity Recognition and Disambiguation, documents can be indexed by invariant concept IDs.
Traditional Indexing:
"The patient was diagnosed with a neoplasm." ➔ [patient, diagnosed, neoplasm]
Entity-Centric Indexing:
"The patient was diagnosed with a neoplasm." ➔ [patient, diagnosed, Entity_C0027651]
This strategy drastically improves precision and recall in specialized domains. For instance, Elena Tutubalina et al. demonstrate in A Comprehensive Evaluation of Biomedical Entity-centric Search that indexing documents by unique biomedical entity identifiers allows systems to handle complex biological hierarchies and nomenclature variations that defeat keyword search.
In the legal domain, Valerio Bellandi et al. show in A Service Infrastructure for Management of Legal Documents that entity-centric modeling allows systems to untangle homonyms and track specific individuals across thousands of unstructured court rulings.
When search engines prioritize these structured entity fields over raw words, the quality of the results improves significantly. Emrah Inan et al. demonstrate this empirical advantage in HSEarch: Semantic Search System for Workplace Accident Reports, showing that entity-based indexing yields a substantially higher normalized discounted cumulative gain (nDCG) than traditional word-based indexing.
Entities as High-Signal Information Anchors
Why do entities provide a stronger retrieval signal than standard text terms? The answer lies in their semantic density and how they partition information.
In any corpus, words follow a Zipfian distribution: a few words appear everywhere but mean very little, while descriptive words are rare and easily missed. Named entities, however, carry high semantic value regardless of their raw frequency. Bingqing Wang et al. leverage this property in Relevance Feedback Based on Constrained Clustering: FDU at TREC 09, extracting people, locations, and organizations independently so they are not overshadowed by statistically dominant but less descriptive terms during query expansion.
Furthermore, treating multi-word entities as single semantic units prevents the false matches that occur when search engines evaluate query terms as independent tokens, a technique validated by Yuanhai Xue et al. in ICTNET at Web Track 2013.
Query: "White House press secretary"
Token-based match: "...met at a white house in the suburbs to discuss the secretary..." (Irrelevant)
Entity-based match: "...the [Entity_White_House] [Entity_Press_Secretary] announced..." (Relevant)
This high information density makes entities exceptional tools for rapidly narrowing down massive search spaces. In high-recall tasks like legal discovery or systematic reviews, finding the last few relevant documents is notoriously difficult. Jie Zou et al. address this in Technology Assisted Reviews: Finding the Last Few Relevant Documents by Asking Yes/No Questions to Reviewers by using entities as information anchors. Instead of asking users to label entire documents, the system asks simple yes/no questions about specific entities. Because entities are highly localized nodes of meaning, a single answer effectively partitions the probability mass of the remaining document collection, shrinking the search space far faster than document-level active learning.
Bridging Queries and Knowledge Graphs
When a query is grounded in specific entities, the retrieval system can move beyond surface text similarity and tap into the relational geometry of a Knowledge Graph (KG).
Zhenghao Liu et al. explore this synergy in Explore Entity Embedding Effectiveness in Entity Retrieval, demonstrating that incorporating low-dimensional KG embeddings (such as TransE) into ad-hoc entity retrieval provides a significant performance boost over purely text-based models. The structured relationships in the graph allow the system to understand the underlying intent of queries that require deep semantic reasoning.
Entity grounding also introduces robust invariance to string matching. In open-domain question answering, models are frequently penalized for generating correct answers that do not match a single gold standard string (e.g., "Tim Cook" vs. "Timothy Donald Cook"). Chenglei Si et al. show in What's in a Name? Answer Equivalence For Open-Domain Question Answering that treating answer verification as an entity-linking problem—by mining entity aliases from a knowledge base—resolves this brittleness.
This structural bridging is equally effective for complex, hierarchical queries. Xinshi Lin et al. demonstrate in CUIS Team at TREC 2018 CAR Track that integrating entity mentions and overlaps through smoothed language models provides a highly reliable relevance signal, successfully connecting complex query headings to knowledge-rich paragraphs.
Two-Stage Discovery and Targeted Updates
Entity-centric retrieval often operates as a two-stage process: using documents as intermediate nodes to discover entities, or conversely, using entities to isolate precise document context.
In entity discovery tasks, the goal is to find entities related to a query rather than documents. Zhanyi Wang et al. (BUPT at TREC 2009: Entity Track) and Qi Li (Searching for Entities When Retrieval Meets Extraction) frame this by treating documents as a "bridge." The system first retrieves relevant supporting documents using standard IR methods, and then extracts and ranks the target entities contained within those documents.
Query ➔ [Stage 1: Retrieve Supporting Documents] ➔ [Stage 2: Extract & Rank Entities] ➔ Target Entity
This two-stage relationship is highly symmetrical. When verifying claims or updating knowledge bases, identifying core entities within a claim serves as a powerful filter to isolate evidence. Wenpeng Yin et al. demonstrate in TwoWingOS: A Two-Wing Optimization Strategy for Evidential Claim Verification that prioritizing entity mentions within claims narrows down massive corpora to a tiny set of relevant evidence pages, significantly outperforming standard keyword similarity.
This precision becomes critical when modifying or updating knowledge within large language models. In document-level model editing, standard vector-based similarity searches often fail because they fetch broad, irrelevant context when multiple similar facts coexist in a document. Zeng Li et al. show in DocMEdit: Towards Document-Level Model Editing that entity-centric retrieval mechanisms drastically outperform standard dense embeddings. By using explicit entity identifiers to fetch factual updates, models isolate the exact information needed to update a document without introducing corruption or being overwhelmed by surrounding context.
Similarly, for local model disambiguation, Kosuke Nishida et al. demonstrate in Self-Adaptive Named Entity Recognition by Retrieving Unstructured Knowledge that formulating queries from specific, uncertain entity spans—rather than the entire input sequence—fetches highly targeted context from external knowledge bases, allowing the model to self-correct its predictions at inference time.
Key Foundational Work
THUIR at TREC 2005: Enterprise Track (2005) — contributed 1 concept elements to this direction.
Language Models for Expert Finding--UIUC TREC 2006 Enterprise Track Experiments (2006) — contributed 1 concept elements to this direction.