Quotegraph: Quotation-Derived Social Network
- Quotegraph is a large-scale, quotation-derived social network linking 528K public figures via speaker-to-mentioned relationships in English news.
- Its construction uses tokenization and heuristic entity linking to extract 8.63M directed edges with contextual metadata from Quotebank.
- The dataset enables computational social science studies, including analyses of gender bias, political discourse, and public reference patterns.
Quotegraph is a large-scale directed social network extracted from speaker-attributed quotations in English news articles published between 2008 and 2020. Its basic relation is speaker-to-mentioned-person: when a quoted speaker mentions another person inside the quotation, Quotegraph creates a directed edge from the speaker to that mentioned person. In the released network, this yields 528 thousand unique nodes and 8.63 million directed edges. The nodes are linked to Wikidata, and the quotation-derived relations retain contextual information from the underlying news corpus, making Quotegraph both a network resource and a quotation-centered data model for computational social science (Čuljak et al., 23 Jul 2025).
1. Definition and corpus foundations
Quotegraph is derived from Quotebank, a large corpus of speaker-attributed quotations from English news. In the Quotegraph paper, Quotebank is described as containing 178 million quotations from 162 million English news articles published between 2008 and 2020. A related Quotebank systems paper presents the broader resource as a corpus with article-level and quote-level views, including quote mentions, ranked speaker-candidate lists, context windows around quote mentions, and canonicalized quote aggregates across articles (Čuljak et al., 23 Jul 2025, Vuković et al., 2022).
This provenance matters because Quotegraph is not a generic social graph extracted from full articles. It is specifically a graph of public reference as expressed in quoted speech. Its edges do not represent friendship, following, reply structure, or hyperlinking. They represent a narrower but analytically distinct relation: who publicly talks about whom, as captured by attributed quotations in news. The project is explicitly positioned as complementary to online social networks, which the authors describe as useful but shaped by platform constraints, moderation, editable content, and misleading AI-generated content (Čuljak et al., 23 Jul 2025).
A common misunderstanding is to read Quotegraph as a quote-propagation graph. The graph itself is not organized around quote-to-quote diffusion. Its primary nodes are person entities, and its primary edges are speaker-to-mentioned-person relations. Quotebank’s article-level and quote-level structures can support downstream studies of quote propagation, quote clustering, or quote-network construction, but those are distinct from the social relation encoded by Quotegraph (Vuković et al., 2022).
2. Construction pipeline
The construction pipeline begins with Quotebank’s article-centric representation. The Quotegraph paper states that article contents are tokenized as lists of Stanford CoreNLP tokens, and that the article-centric data includes quotation offsets and spans of named entity mentions. Because quotation end offsets are not directly provided, the authors reconstruct quote spans by starting from the quotation start token and scanning forward until the closing quotation mark. Since Quotebank extracts the innermost quote in nested cases, the first closing quotation mark is taken as the quote end (Čuljak et al., 23 Jul 2025).
Identifying person mentions inside quotations is handled conservatively. The paper reports that naïve span-overlap checks between quotation spans and mention spans were unreliable because the spans were “not always consistent.” The implemented solution concatenates the quotation tokens into a string , concatenates each named-entity mention into a string , and treats the mention as internal to the quote when is a substring of . The resulting person mentions are then linked to Wikidata entities using lightweight heuristics from quotebank-toolkit rather than a heavier entity-linking system. The paper motivates this choice by estimating that linking all 573M entity mentions in Quotebank with ReFinED would take approximately 95 days on an academic budget, whereas the heuristic linker processes all mentions in approximately two hours and requires no training data (Čuljak et al., 23 Jul 2025).
Several filtering rules are applied before graph construction. Quotations with fewer than case-folded unique words, excluding punctuation, are removed. Similar quotations are grouped when they share a case-folded substring of at least words, again excluding punctuation, and the longest quotation in the group is retained. To reduce spurious person mentions, the system excludes one-character tokens, tokens with no alphabetical characters, and stopwords as possible mentions. Self-loops are removed because they are treated as likely artifacts of attribution error (Čuljak et al., 23 Jul 2025).
Speaker and mention identities are aggregated across contexts. For speaker attribution, local attribution probabilities are first mapped from names to entities and then summed over all quotation contexts; the quotation is assigned to the entity with the highest global attribution probability. For mentions, the most common mention is selected as the target node, and when a quotation contains multiple mentions, the most common set of Wikidata entities is chosen and a distinct edge is created for each entity in that set. This produces the final speaker-to-mentioned-person graph (Čuljak et al., 23 Jul 2025).
3. Data model and contextual enrichment
Quotegraph nodes are person entities identified by Wikidata QIDs. The fundamental edge unit is explicitly defined by the triplet
This means the graph is not merely a collapsed adjacency matrix over person pairs. At its finest granularity, each relation is anchored in a particular quotation event, and repeated mentions across time or articles can be aggregated afterward (Čuljak et al., 23 Jul 2025).
The graph is context-enriched because it inherits metadata from Quotebank. Relations can be traced back to the quotation, the article context in which the quotation appeared, URLs of articles in which the quotation appeared, and the earliest appearance date. This design gives Quotegraph a richer interpretive layer than a plain edge list, since relations remain connected to the linguistic and journalistic contexts from which they were derived (Čuljak et al., 23 Jul 2025).
Wikidata linkage adds structured person attributes. The paper explicitly discusses nationality, gender, date of birth, political party affiliation, and occupation. If multiple birth dates are listed, which occurs for only 0.1% of entities in the extracted networks, the first listed birth date is used. All listed nationalities are extracted, but countries that no longer exist, such as the Soviet Union and the Socialist Federal Republic of Yugoslavia, are excluded. Gender is collapsed into three categories: female, male, and other, with “other” covering non-binary gender or multiple listed genders. Political party affiliation is made time-sensitive using the quotation date together with Wikidata start and end dates; if no temporal bounds are available, the last listed party is used, and if only coarse dates are available, the earliest compatible date is assumed. Occupations are mapped through the Wikidata subclass hierarchy into the broad domains art, sport, politics, and other (Čuljak et al., 23 Jul 2025).
This metadata design makes Quotegraph simultaneously a quotation-derived social network and an entity-enriched research dataset. A plausible implication is that the graph is especially suited to studies of public discourse stratified by demographic or institutional attributes rather than only by topology.
4. Structural properties and descriptive profile
The released graph has the following reported structural characteristics (Čuljak et al., 23 Jul 2025).
| Property | Value |
|---|---|
| Nodes | 528k |
| Directed edges | 8.63M |
| Mean degree | 32.77 |
| Weakly connected components | 5651 |
| Nodes in largest connected component | 97.53% |
| Degree assortativity | 0.034 |
| Global clustering coefficient | 0.265 |
| Assortative mixing by nationality | 0.326 |
| Assortative mixing by occupation | 0.283 |
| Assortative mixing by gender | 0.052 |
The paper describes the indegree and outdegree distributions as heavy-tailed and interprets the network as exhibiting properties typical of large real-world social networks. Most nodes lie in one giant connected component, degree assortativity is positive but low, and clustering is substantial. Attribute assortativity is strongest for nationality and occupation, whereas gender and political party affiliation are much weaker mixing dimensions. Party assortativity is reported separately as 0.024 for the USA, 0.121 for the UK, and 0.072 for India (Čuljak et al., 23 Jul 2025).
The descriptive profile is computed in degree-weighted form, so the resulting distributions characterize interaction mass rather than merely the count of persons. On that basis, actors from the USA account for 48.4% of interactions. Sportsmen are the most represented occupational category, but the most central nodes by PageRank are politicians. The age distribution has two peaks, around 28 and after 47, which the paper interprets as reflecting younger sportspeople and older political actors. Men appear in approximately 87% of interactions, which the authors treat as evidence of strong gender imbalance in news-mediated discourse (Čuljak et al., 23 Jul 2025).
5. Research uses and substantive findings
The paper presents Quotegraph as a resource for computational social science. Proposed use cases include studying the evolution of political polarization among political actors, distinguishing periods of person-centered from issue-centered discourse, and analyzing when politicians mention persons from other occupations and vice versa. Because the graph spans more than a decade and preserves quotation-level context, it can support both longitudinal and attribute-aware analyses (Čuljak et al., 23 Jul 2025).
A concrete example in the paper is nominal gender bias. Using Quotegraph together with Wikidata metadata, the authors study whether a mentioned person is referred to by first name, last name, or full name, while controlling for occupation, nationality, age, prominence approximated through PageRank, name length, and name frequency. Their reported finding is that women are approximately twice as likely to be referred to by their first name as men. The paper interprets this as evidence that nominal gender bias is visible not only in casual discourse but also in professional public communication as captured by quoted news speech (Čuljak et al., 23 Jul 2025).
This application is methodologically important because it shows that Quotegraph is not only useful for network topology. Its combination of person entities, quote-local mention forms, and contextual metadata supports sociolinguistic analysis over networked discourse. In that sense, Quotegraph occupies a middle ground between text corpora and social networks.
6. Scope, limitations, and relation to adjacent quotation-network research
Quotegraph is intentionally narrow in what it represents. It does not capture all real-world social relations; it records only those cases in which a person is quoted in news, the quotation is attributed, another person is mentioned inside the quotation, and both speaker and mention are linked successfully. The paper explicitly acknowledges that some interactions in the giant connected component are likely artifacts of imperfect speaker attribution and entity disambiguation. As upstream evidence, it cites Quotebank/Quobert as correctly attributing 86.9% of quotations in Quotebank experiments, but it does not provide a dedicated intrinsic precision/recall evaluation for final Quotegraph edges (Čuljak et al., 23 Jul 2025).
The corpus is also bounded by English-language news from 2008 to 2020. Coverage therefore inherits editorial priorities, outlet selection effects, and the geopolitical skew of English-language reporting. The strong presence of US actors and male actors is not only a property of public discourse; it is also a property of the news ecosystem through which Quotegraph is mediated (Čuljak et al., 23 Jul 2025).
At the same time, the pipeline is described as language-agnostic because it relies on quotation extraction, tokenization, heuristic entity linking, and Wikidata identifiers rather than language-specific supervised models. This suggests portability to non-English corpora, but that portability is methodological rather than empirically demonstrated in the released dataset (Čuljak et al., 23 Jul 2025).
In a broader research context, Quotegraph can be read alongside two adjacent lines of work. First, the Quotebank systems paper shows that article-level and quote-level representations can support quote propagation and clustering, even though that paper does not formalize a graph (Vuković et al., 2022). Second, the quoter model proposes a paradigmatic model of the social flow of written information in which nodes on a graph generate symbolic time series and copy short text segments from one another (Bagrow et al., 2017). This suggests a useful conceptual distinction: Quotegraph operationalizes observable public reference relations in news quotations, whereas models of textual flow treat quotation as a mechanism of information propagation. The two perspectives are complementary rather than equivalent.
Overall, Quotegraph is best understood as a quotation-derived, Wikidata-enriched social network of public figures in news. Its central innovation is not generic graph extraction, but the conversion of attributed quotations into a person-to-person relation with retained textual context, scalable entity linking, and demographic enrichment.
