Meta WikiExpert: Semantic Linking in Wikipedia

• Meta WikiExpert is a system that leverages a third-party taxonomy to connect encyclopedic content with external expert profiles without modifying Wikipedia articles.
• It utilizes a multi-stage pipeline—including parsing, taxonomy indexing, and SPARQL querying—to associate Wikipedia pages with expert data from sources like Inria reports.
• The approach supports diverse deployments, extending to scientific abstract wikification, mathematical formula linking, and multilingual citation-grounded applications.

Meta WikiExpert is a Wikipedia-centered system for connecting encyclopedic pages to external knowledge resources through a shared semantic layer, so that ordinary page navigation can surface domain-specific expertise or grounded explanatory material without altering core Wikipedia content (Grefenstette et al., 2015). In its foundational form, the system re-indexes both Wikipedia and a non-Wikipedia corpus under a third-party taxonomy independent from Wikipedia’s category hierarchy, stores the resulting associations in a relational database or triple store, and exposes them through a SPARQL endpoint consumed by a Wikipedia gadget (Grefenstette et al., 2015). Related work extends the same general design space to scientific abstract wikification through metapaths (Jana et al., 2017), mathematical formula linking to Wikidata (Scharpf et al., 2021), and multilingual citation-grounded question answering and report generation using MegaWika (Barham et al., 2023). This suggests a broader class of “Meta WikiExpert” deployments in which Wikipedia functions simultaneously as interface, pivot resource, and semantic access point.

1. Concept and scope

The motivating premise is that Wikipedia is widely used for finding general information about a wide variety of topics, but its vocation is not to provide local information such as local scientific experts (Grefenstette et al., 2015). Meta WikiExpert addresses that gap by connecting Wikipedia pages to external resources through a third-party taxonomy, rather than by modifying article prose or relying on Wikipedia’s native category system. In the original case study, the non-Wikipedia corpus consists of local expert descriptions drawn from Inria annual reports, and the shared semantic pivot is ACM CCS in SKOS (Grefenstette et al., 2015).

A central design property is separation of concerns. Wikipedia remains the front-end navigation environment; the external corpus remains external; and the semantic linkage is mediated by an ontology-like taxonomy. The technique is presented as generic: any third party taxonomy can be used to connect Wikipedia to any non-Wikipedia data source (Grefenstette et al., 2015). A plausible implication is that the defining characteristic of Meta WikiExpert is not the specific expert-finding use case, but the use of Wikipedia as a semantic access layer over externally indexed resources.

2. System architecture and user interaction

The Meta WikiExpert system comprises six main components: Wikipedia Parser, Taxonomy Indexer, Expert Profile Indexer, Relational Database (or Triple Store), SPARQL Endpoint, and Wikipedia Gadget (ResourceLoader JavaScript) (Grefenstette et al., 2015).

Component	Role
Wikipedia Parser	Extracts article IDs, titles, and body text from a recent Wikipedia XML dump
Taxonomy Indexer	Matches controlled-vocabulary terms to Wikipedia and expert text spans
Expert Profile Indexer	Applies the same NLP pipeline to local expert profiles
Relational Database or Triple Store	Stores concept associations
SPARQL Endpoint	Exposes the associations for querying
Wikipedia Gadget	Queries the endpoint on page view and displays matched experts

The data flow is explicitly staged. A scheduled job pulls a recent Wikipedia XML dump; the Wikipedia Parser extracts article IDs, titles, and body text, then tokenizes, lowercases, and splits into sentences. The same NLP pipeline is applied to local expert profiles, such as Inria annual report HTML pages, by the Expert Profile Indexer. The Taxonomy Indexer then loads the third-party taxonomy and, for each sentence in Wikipedia or expert text, matches controlled-vocabulary terms to text spans, producing document–concept associations. These associations are stored in relational tables or rewritten as RDF triples and made available through a SPARQL endpoint (Grefenstette et al., 2015).

On the client side, the user-installed gadget intercepts page view events in the browser, extracts the current article URI, and asynchronously queries the SPARQL endpoint to fetch all experts linked via shared taxonomy concepts. The gadget adds an additional tab, such as “Inria,” to the page; when clicked, it displays the matched experts and topics. The interface inherits the look-and-feel of standard Wikipedia tabs through the CSS class vector-menu, and the dropdown organizes results by concept headings with linked team names beneath them (Grefenstette et al., 2015).

This architecture preserves Wikipedia as the interaction surface while externalizing indexing, storage, and ranking logic. An immediate consequence is that deployment does not require direct modification of Wikipedia content, only client-side augmentation plus a queryable semantic backend.

3. Pivot ontology, scoring, and indexing

The pivot ontology is defined as a set of concepts $C = \{c_1, c_2, \dots, c_n\}$, where each concept has skos:prefLabel, skos:altLabel, and skos:broader relations, together with a set of literal terms

$T = \{t_1, t_2, \dots, t_m\}$

drawn from preferred labels, synonyms, and Wikipedia redirects (Grefenstette et al., 2015). This construction makes the taxonomy a common indexing vocabulary over otherwise heterogeneous corpora.

For each term $t_i \in T$ and each text segment $W_j$—either a Wikipedia article or an expert report—the system computes a normalized tf–idf score:

$$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$

Here, $tf(t_i, W_j)$ is the number of times $t_i$ occurs in $W_j$, $df(t_i)$ is the number of documents in which $t_i$ occurs, $T = \{t_1, t_2, \dots, t_m\}$0 is the total number of documents across Wikipedia and expert profiles, and $T = \{t_1, t_2, \dots, t_m\}$1 is the length of $T = \{t_1, t_2, \dots, t_m\}$2 in tokens (Grefenstette et al., 2015).

The disambiguation and pruning heuristics are deliberately simple. The system eliminates any concept whose terms appear in more than 10 000 Wikipedia articles, treating such concepts as too general. If multiple concepts match the same span, it prefers the more specific one, meaning the concept deeper in the hierarchy. Matching uses exact string match on lowercased tokens and requires term boundaries or stopword-delimited tokens. Wikipedia redirect titles are added to the term inventory to boost recall (Grefenstette et al., 2015).

The indexing logic for Wikipedia and expert profiles is parallel. For each sentence, the system detects terms from $T = \{t_1, t_2, \dots, t_m\}$3, maps them to concepts, and emits (page_id, concept_id) or (expert_id, concept_id) pairs when the sentence-level score exceeds threshold $T = \{t_1, t_2, \dots, t_m\}$4 or $T = \{t_1, t_2, \dots, t_m\}$5. The thresholds are tuned to balance precision and recall, with the details giving $T = \{t_1, t_2, \dots, t_m\}$6 as an example (Grefenstette et al., 2015).

The resulting associations support direct graph-style retrieval through SPARQL. The example query pattern links a currently viewed Wikipedia page to teams sharing one of its concepts: $tf(t_i, W_j)$2 This makes the pivot ontology not merely an indexing device, but the formal interface by which Wikipedia pages become retrieval keys for external expertise (Grefenstette et al., 2015).

4. Scientific and mathematical semantic linking

A related line of work studies wikification in scientific abstracts, where the task is to connect technical mentions in short, dense research summaries to Wikipedia pages (Jana et al., 2017). WikiM treats this as a two-stage problem: mention extraction followed by entity linking. Mention extraction begins with tokenization and noise removal, then applies POS tagging to identify adjective- or noun-laden $T = \{t_1, t_2, \dots, t_m\}$7-grams up to length three. Overlapping spans are resolved in favor of the longer phrase, and surviving fragments are retained only if a Wikipedia page or disambiguation entry exists. Ranking is then based on a tf–idf score computed strictly on the 2013 ACL Anthology Network corpus:

$T = \{t_1, t_2, \dots, t_m\}$8

The system selects 4 mentions for abstracts with at most two sentences, 8 mentions for abstracts with three to four sentences, and 12 mentions otherwise (Jana et al., 2017).

For disambiguation, WikiM constructs a heterogeneous graph with paper nodes, author nodes, and year nodes, and defines author, citation, reference, and year-restricted metapaths. When the top two candidate Wikipedia entities for a mention fall within a small cosine-similarity margin $T = \{t_1, t_2, \dots, t_m\}$9, the system treats the case as ambiguous and expands the document context through reachable papers whose abstracts are sufficiently similar. For non-acronyms, candidate entities are scored by the overlap between $t_i \in T$0-grams in the Wikipedia summary and those in the union of the abstract and metapath context. For acronyms, a weighted interpolation is used:

$t_i \in T$1

with $t_i \in T$2 (Jana et al., 2017).

On the ACL Anthology Network, WikiM reports 72.42% precision and 72.1% recall for mention extraction, with $t_i \in T$3, and 73.8% precision for entity linking on correctly extracted mentions. The paper also reports that adding author metapaths raises linking precision from near 63% to 71.4%, while the full citation + reference + author metapaths reach 73%, and the five-year restriction adds a further 0.8% gain (Jana et al., 2017). In a Meta WikiExpert context, this establishes that scholarly network structure can compensate for the weak local context of short technical texts.

A mathematical counterpart is provided by AnnoMathTeX, which links mathematical expressions in Wikipedia or LaTeX-based STEM documents to semantic concepts in Wikidata (Scharpf et al., 2021). Its three-step data-refinement pipeline is: “Mine and recommend,” “Seed Wikidata,” and “Publish links.” The system parses Wikitext or raw LaTeX, extracts formulae and constituent identifiers, builds an in-memory model of formula objects and identifier objects, and lets editors annotate either entire formulae or symbols via popup interfaces. Global annotations for a symbol propagate automatically to its other occurrences (Scharpf et al., 2021).

Recommendation sources are heterogeneous. Identifier suggestions come from arXiv corpus frequencies over 60 M mathematical expressions in the NTCIR arXiv benchmark, Wikipedia definitions, Wikidata quantity-symbol property P416, a local word window of $t_i \in T$4 words, and user-typed memory. Formula suggestions come from fuzzy matching on Wikidata P2534 (“defining formula”), parts overlap through annotated identifiers, formula-concept memory, a $t_i \in T$5 word window, and user input. The quality of a source is measured by the rank at which an editor accepted its suggestion, using

$t_i \in T$6

with $t_i \in T$7 and $t_i \in T$8 if the editor accepted the suggestion at rank $t_i \in T$9, else $W_j$0 (Scharpf et al., 2021).

In a 25-article experiment on English Wikipedia physics articles, AnnoMathTeX reports annotation speed-ups from 6.3 s to 2.6 s for identifiers, or 2.4× faster, and from 4.0 s to 2.8 s for formulae, or 1.4× faster. After seeding, approximately 80% of identifiers and approximately 60% of formulae had a valid Wikidata QID. Within a one-month window, 12% of formula links in Wikipedia articles were reverted or commented on, and 33% of seeded Wikidata items were modified (Scharpf et al., 2021). This places mathematical semantic linking within the same general ecosystem as Meta WikiExpert, but with Wikidata rather than a third-party taxonomy serving as the semantic target.

5. Multilingual citation-grounded expansion

MegaWika extends the Meta WikiExpert design space from ontology-based expert lookup to multilingual citation-grounded report generation and question answering (Barham et al., 2023). The dataset contains 13 million Wikipedia articles in 50 diverse languages together with 71 million referenced source materials. The English portion alone includes 1.20 M articles, 3.45 M cited source documents, 9.1 M paragraphs, and 45.7 M sentences; the median over 50 languages is 64 k articles, 345 k sources, 490 k paragraphs, and 2.4 M sentences (Barham et al., 2023).

Preprocessing proceeds through HTML-to-plain-text conversion, boilerplate removal, sentence segmentation using a language-specific tokenizer such as ICU rules, citation–paragraph alignment, and bilingual alignment for non-English material. Citation–paragraph alignment assigns each reference marker $W_j$1 to the paragraph maximizing token overlap:

$W_j$2

For bilingual alignment, sentence pairs shorter than 5 tokens or longer than 200 tokens are discarded (Barham et al., 2023).

Semantic annotation is layered on top of this corpus. FrameNet parses are produced over English lead paragraphs with the SEMAFOR parser, assigning each predicate trigger to a frame and role fillers, and these parsed frames are projected to other languages via word alignment. The target article is then segmented into “Topic Blocks” guided by semantic frames, and block summary templates instantiate patterns such as “In YEAR, AGENT [F:Create_…] ORGANIZATION” (Barham et al., 2023).

The question-answering component uses an XLM-R base encoder $W_j$3 with a span-prediction head $W_j$4:

$W_j$5

The reported training regime is pretraining on Wikipedia-derived MLAQA splits and fine-tuning on MegaWika-QA with 2 M QA pairs, batch size 32, learning rate $W_j$6, 3 epochs, and 10% warmup. On a held-out multilanguage test set, sample results are English $W_j$7, EM = 68.5; Spanish $W_j$8, EM = 64.3; Arabic $W_j$9, EM = 54.9. The ablations –frameline and –cross-align reduce $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$0 by 2.3 and 1.7 points respectively (Barham et al., 2023).

Citation retrieval is handled by a DPR-style dual encoder with question encoder $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$1 and passage encoder $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$2, scored by dot product:

$$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$3

At $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$4, the example retrieval results are English $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$5, $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$6, and Spanish $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$7, $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$8 (Barham et al., 2023).

The explicit integration path for a Meta WikiExpert system is: a user issues query $$\mathrm{score}(W_j, t_i) = \frac{ tf(t_i, W_j)\,\times \log\bigl(\tfrac{N}{df(t_i)}\bigr) }{ \|W_j\| }.$$9 in any of 50 languages; the retriever fetches top-$tf(t_i, W_j)$0 grounded paragraphs across languages; the reader extracts answer spans and associated citations; semantic assembly uses frames to weave the answer into a coherent paragraph; and the final output is a multilingual narrative with in-text citation markers linking back to sources (Barham et al., 2023). This shifts Meta WikiExpert from concept lookup toward grounded explanatory synthesis.

6. Evaluation, limitations, and open questions

In the original local-expert deployment, the system indexed 129 499 Wikipedia articles with 1 049 distinct ACM concepts and indexed 3 123 Inria web pages from 2014 activity reports (Grefenstette et al., 2015). The paper does not report a formal user study, but it identifies standard IR metrics that would apply, including Precision@k, Recall, and $tf(t_i, W_j)$1 (Grefenstette et al., 2015). The absence of a formal user study is significant because it leaves open whether semantic linkage quality, interface familiarity, and expert usefulness align in realistic navigation settings