Enhanced Entity Page Format

• Enhanced entity page format is a comprehensive approach that presents a canonical entity as a self-contained, richly interlinked, and machine-actionable resource.
• It utilizes interoperable page primitives such as mention spans, template-driven sections, JSON-LD embeddings, and typed relation graphs for improved entity representation.
• The format supports dynamic tasks including retrieval augmentation and efficient summarization through preview tables and agent instructions to enhance both maintenance and search.

An enhanced entity page format is a human- and agent-optimized web or knowledge-base document that presents a single entity as a self-contained, richly interlinked, machine-actionable resource. In the research literature, the notion extends beyond web presentation alone: it also denotes page designs that expose a canonical topic entity, align sentence- and section-level structure to that topic, support template-driven updates from external news, compress large entity graphs into preview tables, summarize RDF-style entity descriptions under explicit salience and diversity criteria, and render typed relation graphs with provenance for exploratory search and profiling (Volpini et al., 11 Mar 2026, Li et al., 2023, Fetahu et al., 2017, Yan et al., 2014, Liu et al., 2019, Amal et al., 2021).

1. Conceptual scope and recurring page primitives

Across the cited work, enhanced entity pages recur as a set of interoperable page primitives rather than as a single fixed schema. These primitives include canonical topic identity from the page URL or title, mention spans with entity IDs and offsets, class-specific section templates, compact previews of entity-type neighborhoods, natural-language summaries derived from structured data, embedded Schema.org JSON-LD, visible dereferenceable links, LLMs.txt-style agent instructions, breadcrumbs, statements tables, sitelinks, and explicit provenance or confidence signals (Li et al., 2023, Fetahu et al., 2017, Yan et al., 2014, Volpini et al., 11 Mar 2026, Amal et al., 2021).

Research line	Page primitive	Representative source
Topic-aware KEPLM pretraining	topic entity, mention spans, injection targets	(Li et al., 2023)
Template-driven page upkeep	article–entity placement, article–section placement, class templates	(Fetahu et al., 2017)
Schema exploration	key/non-key preview tables with size and distance constraints	(Yan et al., 2014)
RAG-oriented linked data pages	JSON-LD, dereferenceable URIs, breadcrumbs, agent instructions	(Volpini et al., 11 Mar 2026)
Summarization and profiling	top-$k$ facts, typed entity–relation graph, provenance	(Liu et al., 2019, Amal et al., 2021)

This recurrent design vocabulary indicates that enhancement is not reducible to adding markup. It involves exposing entity identity and relation structure in forms consumable by both humans and automated systems. A plausible implication is that the “page” serves simultaneously as presentation layer, training substrate, retrieval unit, and maintenance target.

2. Topic awareness as a representational principle

Wikipedia has a topic-centered layout in which the page URL and title uniquely identify a topic entity $e_t$, while hyperlinks inside sentences identify mentioned entities $E=\{e_i\}$. KEPLET formalizes each pretraining sentence as $\langle e_t, W, E\rangle$ and argues that conventional knowledge-enhanced pretrained LLMs neglect this page-level topic signal, producing insufficient entity interaction and biased relation word semantics (Li et al., 2023). The paper’s empirical intuition is that topic entities rarely appear explicitly in their own page sentences: only about $6\%$ of top-500K popular pages mention their topic entities in the sentence text. As a result, a naive positional treatment of $e_t$ is impractical.

KEPLET introduces topic entity awareness through three coupled mechanisms. First, it learns an injection gate at each token or entity position,

$$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$

so that topic information is inserted where it is most useful rather than where it is merely available. Second, it fuses topic features into token and mention representations through an adapter, with attention fusion preferred in the reported experiments. Third, it adds topic-entity-aware contrastive learning, aligning sentence-level $[\mathrm{CLS}]$ representations and topic entity vectors within the same page-topic pool. The full pretraining objective augments the base KEPLM loss with topic contrastive supervision,

$$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$

The model is trained end-to-end on December 2018 English Wikipedia, using hyperlinks as mention spans and page identity as topic identity. The topic module is interleaved between transformer layers, with no parameter sharing across layers; empirically, lower-layer insertion at $l=1$ performs best, and one fusion layer outperforms two or three. At fine-tuning time, the topic fusion module is discarded, while the improved token and entity parameters are retained (Li et al., 2023).

The reported gains are task-specific but consistent across entity-centric benchmarks. For LUKE-base with attention fusion, Open Entity micro-F1 rises from $e_t$0 to $e_t$1, TACRED micro-F1 from $e_t$2 to $e_t$3, SQuAD F1 from $e_t$4 to $e_t$5, and CoNLL-2003 F1 from $e_t$6 to $e_t$7. For ERNIE, TACRED improves from $e_t$8 to $e_t$9. Attention fusion also outperforms concatenation on the cited comparisons, such as Open Entity $E=\{e_i\}$0 versus $E=\{e_i\}$1 and TACRED $E=\{e_i\}$2 versus $E=\{e_i\}$3 (Li et al., 2023).

The canonical example is the sentence “She released Crazy in Love and Baby Boy.” On the Beyoncé page, KEPLET identifies the topic entity as Beyoncé and tends to inject topic information into the pronoun “She” and the ambiguous mention spans. This reorients the semantics of “released” toward a topic–mention relation rather than a merely local token–mention pattern. The larger significance is that enhanced entity pages can expose topic identity explicitly enough for representation learning without requiring that the topic entity be restated in every sentence.

3. Template-driven structure, timeliness, and page maintenance

A second line of work treats enhanced entity pages as update targets whose sections should be populated with timely, relevant references. “Automated News Suggestions for Populating Wikipedia Entity Pages” frames this as a two-stage supervised problem: Article–Entity Placement (AEP), which decides whether a news article should be suggested to an entity page, and Article–Section Placement (ASP), which selects the exact page section for the citation (Fetahu et al., 2017). The formal tasks are

$E=\{e_i\}$4

The motivation is both documentary and structural. The paper notes that as much as $E=\{e_i\}$5 of Wikipedia references are from online news sources, yet many entity pages remain incomplete or lag behind the publication time of relevant news. AEP uses salience, relative authority, and novelty to filter relevant article–entity pairs, while ASP maps accepted articles to class-appropriate sections using topic, lexical, syntactic, and entity-type signals. Templates $E=\{e_i\}$6 are learned by grouping entity pages by class and clustering section text with x-means over tf–idf vectors using cosine similarity. Supported classes in evaluation cover 27 DBpedia classes, including Person subclasses, Organization, Location, Event, and Creative Work subclasses (Fetahu et al., 2017).

The page-level implications are direct. Completeness follows because ASP can propose missing but class-typical sections, such as “Accidents/Incidents” for airlines. Timeliness follows because AEP monitors news streams and privileges salient, novel content. Structural consistency follows because the same class-based template constrains citation placement across entities of the same type. In production terms, the format becomes template-driven rather than purely editorially ad hoc.

The reported results are strong for both stages. On AEP, the system achieves up to $E=\{e_i\}$7 precision; average performance across years is approximately $E=\{e_i\}$8, $E=\{e_i\}$9, and $\langle e_t, W, E\rangle$0, with Cohen’s $\langle e_t, W, E\rangle$1. Relative entity frequency $\langle e_t, W, E\rangle$2 is identified as the strongest single contributor, while authority and novelty features add about $\langle e_t, W, E\rangle$3–$\langle e_t, W, E\rangle$4 precision on top of salience signals. On ASP, average performance is approximately $\langle e_t, W, E\rangle$5, $\langle e_t, W, E\rangle$6, and $\langle e_t, W, E\rangle$7, with $\langle e_t, W, E\rangle$8; Person–Politics and Person–Entertainment reach about $\langle e_t, W, E\rangle$9–$6\%$0 F1, Organization about $6\%$1, and Location about $6\%$2 (Fetahu et al., 2017).

This work shifts the meaning of an enhanced entity page from “richly annotated” to “maintainable under typed structural priors.” A plausible implication is that section schemas are not merely presentational conventions but decision surfaces for automated curation.

4. Compression of entity knowledge: preview tables and extractive summaries

Large entity descriptions and heterogeneous knowledge graphs create an opposing problem: excessive structural richness. Two bodies of work address this by selecting compact subsets of facts and relations for early-stage understanding (Yan et al., 2014, Liu et al., 2019). In the preview-table formulation, an entity graph is a directed multigraph $6\%$3 and the schema graph is $6\%$4. A preview table has a key attribute corresponding to an entity type $6\%$5 and non-key attributes corresponding to relationship types incident on $6\%$6. A preview is a set of $6\%$7 such tables with distinct keys.

The scoring model is explicit:

$6\%$8

Key attributes can be scored by coverage or a random-walk stationary distribution on the schema graph, while non-key attributes can be scored by coverage or entropy. The optimization problem is then to find the highest-scoring preview under a size constraint $6\%$9 and, optionally, a distance constraint between preview tables. Tight previews require pairwise distances $e_t$0; diverse previews require pairwise distances $e_t$1. Both tight and diverse preview discovery are NP-hard under the corresponding distance constraints, leading to a dynamic-programming algorithm for concise previews and an Apriori-style algorithm for tight or diverse previews (Yan et al., 2014).

The preview-table results indicate that the scoring measures are effective and the algorithms efficient. Coverage and random-walk scores achieve Precision@K near optimal in four of five evaluated domains; non-key scoring by coverage or entropy yields mean reciprocal rank above $e_t$2 in all domains except film, and user studies with 84 CS graduate students show that tight previews are often the fastest representation for test questions, while graph and schema-summary interfaces can feel more complete even when objective accuracy is lower (Yan et al., 2014).

Entity summarization research addresses a finer granularity: the selection of up to $e_t$3 triples from an entity description $e_t$4. The standard extractive objective is

$e_t$5

The 2019 survey classifies methods by frequency and centrality, informativeness, and diversity and coverage, with domain knowledge, context awareness, and personalization as task-specific extensions (Liu et al., 2019). Existing methods are reported to be mainly unsupervised and to combine technical features through random surfer models, similarity-based grouping, MMR-like re-ranking, or combinatorial optimization; a few deep learning methods appear in more recent work. This taxonomy makes enhanced entity pages interpretable as controlled summaries rather than unfiltered fact dumps.

Taken together, preview tables and entity summaries define two complementary compression regimes. The first summarizes schema-level neighborhoods; the second summarizes fact-level descriptions. This suggests that a mature entity page format can expose both: a compact view of what kinds of things relate to the entity, and a compact view of which concrete facts are most salient.

5. Linked data pages as a memory layer for RAG and agentic retrieval

The most explicit contemporary formulation defines the enhanced entity page as a retrieval substrate. In “Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval,” the page includes a natural-language summary generated from structured data, a complete embedded Schema.org JSON-LD block, visible dereferenceable links to related entities, LLMs.txt-style agent instructions, navigational affordances such as type breadcrumbs and sitelinks, and optional neural search endpoint references (Volpini et al., 11 Mar 2026). The Enhanced+ variant adds richer navigational affordances, named linked entities, statements tables, canonical URIs, and a stronger visual hierarchy.

The architectural claim is that linked data turns pages into addressable memory cells. Each entity URI dereferences via content negotiation to HTML, JSON-LD, or Turtle; agents can follow typed links such as schema:about, schema:author, schema:offers, or schema:containedInPlace; and the page itself exposes enough relational structure that both standard and agentic RAG systems can retrieve and traverse entity-centric facts without relying exclusively on flat-text chunks (Volpini et al., 11 Mar 2026).

The reported experiment covers four domains—editorial, legal, travel, and e-commerce—with 158 entities, 349 test queries, 2,443 evaluations, and 2,439 valid judgments. Seven conditions are tested: plain HTML, HTML with JSON-LD, and enhanced entity pages under standard RAG and agentic RAG, plus Enhanced+ under agentic RAG. Retrieval uses Vertex AI Vector Search 2.0 with hybrid search, top-$e_t$6, and gemini-embedding-001; agentic reasoning uses the Google Agent Development Kit with a ReAct-style loop, tools for document search and link following, and max_hops=2 (Volpini et al., 11 Mar 2026).

The quantitative pattern is clear. JSON-LD alone yields only a small lift: C2 versus C1 improves accuracy by $e_t$7 with adjusted $e_t$8 and $e_t$9. By contrast, enhanced pages materially improve both standard and agentic retrieval. Under standard RAG, C3 versus C1 improves accuracy from $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$0 to $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$1 and completeness from $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$2 to $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$3, corresponding to an accuracy gain of about $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$4. Under the full pipeline, C6 versus C1 improves accuracy from $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$5 to $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$6, about $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$7. Enhanced+ attains the highest absolute scores, with accuracy $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$8 and completeness $$g_p^{(i)} = \sigma\!\left(F_p\!\left(\mathbf{h}^{(i)}\right)\right),$$9, although the increment over the base enhanced format is not statistically significant for accuracy ($[\mathrm{CLS}]$0, adjusted $[\mathrm{CLS}]$1, $[\mathrm{CLS}]$2) (Volpini et al., 11 Mar 2026).

The operational explanation is that explicit relational surface area matters. Enhanced+ exposes about $[\mathrm{CLS}]$3 more links than JSON-LD pages ($[\mathrm{CLS}]$4 versus $[\mathrm{CLS}]$5) and about $[\mathrm{CLS}]$6 more than plain HTML ($[\mathrm{CLS}]$7 versus $[\mathrm{CLS}]$8), yet agents follow fewer links ($[\mathrm{CLS}]$9 versus $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$0 in plain HTML). The format therefore increases context density while reducing multi-hop cost. The paper also reports that $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$1 of plain HTML pages and $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$2 of JSON-LD pages exceed the approximately 20k-character ingestion limit used in the flat-text pipeline, with JSON-LD often starting near character $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$3, making truncation a practical issue for HTML+JSON-LD pages that do not materialize salient facts visibly (Volpini et al., 11 Mar 2026).

In this formulation, the enhanced entity page is not merely “machine-readable.” It is optimized for retrieval fidelity, action planning, and low-hop graph traversal.

6. Typed profile graphs, governance, limitations, and future directions

A further extension treats the enhanced entity page as a live profile graph assembled from web evidence. The Person Entity Profiling Framework begins from a real-time person query, retrieves top-ranked web pages, re-ranks them with supervised relevance classification, extracts entities and relations from relevant pages, aggregates them into a typed entity–relation graph, and visualizes the result in an interactive D3.js force-directed interface (Amal et al., 2021). Relevant-page classification uses lexical, page-type, title-relevance, body-relevance, and rank features; Bagging reaches $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$4 with precision $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$5 and recall $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$6 on a 1,277-page mixed-domain dataset, while a DNN with early stopping reaches about $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$7 on a large auto-labeled congress dataset. Relation typing in the scholar domain reaches about $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$8 for Employment and Education with Logistic Regression and about $$\mathcal{L} = \mathcal{L}_{\text{PLM}} + \mathcal{L}_{\text{Aux.}} + \lambda \sum_S \mathcal{L}_S .$$9 for Publications with SVM; the multi-domain PCNN-with-attention setting reports average precision about $l=1$0, recall about $l=1$1, and F1 about $l=1$2 (Amal et al., 2021).

The rendered page is graph-centric. Node shape and color encode type, relation color encodes relation category, edge width or node size encodes mention-count strength, and temporal information is mapped into five recency buckets: 0–2004, 2004–2007, 2008–2011, 2012–2015, and 2016–2019. Clicking an edge reveals a word cloud of contextual evidence. In the reported user study with 64 participants, 40 preferred the graph visualization over ranked lists, 16 liked both, and 8 preferred lists; filter usefulness ratings are $l=1$3 for relation filters, $l=1$4 for node filters, and $l=1$5 for the node-count slider, while accuracy and coverage are rated $l=1$6 (Amal et al., 2021). The framework also surfaces substantial additional related entities beyond homepages: for actors, $l=1$7 were new and $l=1$8 of those were also on Wikipedia; for singers, $l=1$9 new and $e_t$00 also on Wikipedia; for businessmen, $e_t$01 new and $e_t$02 also on Wikipedia; for politicians, $e_t$03 new and $e_t$04 also on Wikipedia (Amal et al., 2021).

Across the broader literature, several constraints recur. KEPLET requires entity-rich corpora with reliable hyperlink mentions and topic identity, depends on high-quality entity linking during pretraining, incurs additional Wikipedia pretraining cost, and does not explicitly consume infobox or section metadata in the current model (Li et al., 2023). Template-based news suggestion can inherit entity-linking errors, can over-standardize section structure, and can bias toward mainstream domains through domain-authority priors (Fetahu et al., 2017). RAG-oriented enhanced pages are evaluated under flat-text ingestion and a single retrieval stack, so structured-data-aware ingestion might alter the comparative effect of JSON-LD (Volpini et al., 11 Mar 2026). Person profiling explicitly lacks a deployed entity-disambiguation module and is vulnerable to ambiguity in common names (Amal et al., 2021). The entity summarization survey identifies long-tail entities, cross-knowledge-graph alignment, multilinguality, temporal dynamics, and robust personalization as open challenges (Liu et al., 2019).

The future directions named in these works are structurally consistent. They include multilingual topic alignment, domain adaptation to specialized corpora, richer page formats with section-level topic relevance and explicit relation-word schemas, event-specific and multilingual section templates, better novelty and deduplication via semantic embeddings, cross-knowledge-graph alignment, temporal event modeling, and stronger integration of graph structure with retrieval and generation (Li et al., 2023, Fetahu et al., 2017, Liu et al., 2019, Volpini et al., 11 Mar 2026). Suggested privacy and ethics considerations in the profiling setting include respecting robots.txt and site terms, providing opt-out mechanisms, labeling confidence and sources, and avoiding inference of sensitive attributes (Amal et al., 2021).

Taken together, these studies define the enhanced entity page format as a convergent architectural pattern: a page centered on a canonical entity, structurally explicit about sections and relations, compact enough to summarize, rich enough to retrieve and traverse, and instrumented enough to support provenance-aware maintenance and learning.