StackOverflow Knowledge Graph - Extra languages

Stack Overflow knowledge graph: Extra languages

This dataset contains Stack Overflow questions whose tags do not match any configured language category in the release's selection registry, together with their answers and comments. It is the extra-languages fallback graph, not a separate programming language.

The RDF export is produced by stack2graph.

Release

The byte count is the sum of the RDF files as stored inside the archive. The size of the downloadable archive also depends on compression and the included metadata.

Contents and selection

The selected language graph is http://stackoverflow.com/extra-languages. Schema statements are stored in http://stackoverflow.com/schema. The package also includes the full set of shared tag-documentation shards, including documentation assigned to other language graphs. The counts below cover every included shard.

Questions with tags for several programming languages can appear in several language packages. Totals from different packages therefore overlap.

Graph model

Each N-Quads statement contains a subject, predicate, object and named graph IRI. The named graph records the language selection that contains the statement. Entities keep the same IRIs across language graphs, so the same question can be queried in several contexts without introducing a separate question identity. Files are storage shards; splitting a graph across files does not create new logical graphs.

The vocabulary uses the namespace http://stackoverflow.com/schema#, abbreviated as so: below. Its RDF/OWL definitions are included in http://stackoverflow.com/schema.

Questions connect to answers, comments, tags and other questions. The relationship table below gives the direction of each predicate and its exported inverse. An accepted answer is also an answer, and acceptance edges are present only when resolved from the source. Comments can also belong to tag documentation posts. The diagram below shows these entities and relationships.

Graph model

The diagram's combined "Other question" node illustrates the possible directions of so:duplicateOf. It does not assert that a pair of questions is marked as a duplicate in both directions. Each arrow requires its own source PostLinks record. Neither so:duplicateOf nor so:linkedTo has an automatically exported inverse; an incoming edge is a separate source relation, not a generated reverse of the outgoing edge. An incoming edge is stored in its source question's language graph, so displaying it can require loading another language package.

Timestamps

Timestamps are datatype properties whose objects are xsd:dateTime literals. Each field is a separate RDF statement. In the diagram, rectangular panels group these properties beside their entities; the panels are not RDF entities. Missing source values produce no statement.

The packaged SVG shows representative panels for the question, an answer and a comment. The same rules apply to accepted answers, answer comments and tag documentation as listed above. For example, a question's creation time is represented as the following triple, within its named graph (the value is illustrative):

@prefix so: <http://stackoverflow.com/schema#> .
@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .

<http://stackoverflow.com/q/123>
    so:creationDate "2024-03-15T10:30:00Z"^^xsd:dateTime .

These fields describe the source snapshot, not a complete revision history. Last activity or last edit dates alone do not establish technical freshness.

Questions, answers and documentation posts carry source post IDs. Available properties include titles, original HTML bodies, scores, aggregated upvote and downvote counts, timestamps and content licenses. Comments carry text, scores, creation dates and licenses. Question-specific properties include source answer, view and favorite counts. Missing source values are omitted. Individual voters and user-profile entities are not represented by this mapping.

The listed inverse relationships are written as triples, so querying them does not require an inference engine. A link or duplicate edge does not imply a reverse edge. The mapper also writes applicable superclass types. Answer status uses classes such as so:AnsweredQuestion, so:UnansweredQuestion and so:QuestionWithAcceptedAnswer; missing edges alone do not establish unanswered status. An accepted answer or a high score does not establish technical correctness. The ontology defines so:possibleBetterAnswer, but this exporter does not generate that assessment.

How the language graphs connect

Language membership is determined from question tags using the release's selection registry. Matching covers configured base tags, version patterns and aliases. Answers and discussion comments inherit their question's graph membership; their bodies are not classified independently. Questions without a registry match are routed to the http://stackoverflow.com/extra-languages graph in the full export. This category is not an additional match for questions already assigned to a registered language.

The additional extra-languages dataset

extra-languages can be published as its own dataset alongside the 39 language datasets. Its graph IRI is http://stackoverflow.com/extra-languages. It contains questions with no match in the release's configured language registry, including their answers, comments and outgoing question links. This can include questions about unconfigured languages and questions without a recognized language tag; it is not a classification into one additional programming language.

The graph also holds tags without recorded graph usage and tag documentation without a resolved owner. Its publication package includes its discussion shards, the schema and all shared documentation shards. Regular language packages include shared documentation from this graph, but not its complete discussion collection. Load the extra-languages dataset from the same export identity to resolve links whose target discussions are in that graph. It complements the configured language selections without assigning those discussions to another language.

There are three ways language graphs share or connect content:

1. The same discussion in several graphs. If a question matches both Python and Java, its question, answer and comment statements are emitted into both language graphs with the same entity IRIs. This overlap follows source tags; it is not a translation or an inferred similarity relationship.
2. Shared tags and documentation. Tags retain their identities across graphs. Their definitions, wikis, excerpts and documentation comments are routed to the graphs of questions using those tags. Tags without recorded usage, and documentation without a resolved tag owner, go to extra-languages.
3. Links between different questions. so:linkedTo and so:duplicateOf point from a question to another question's IRI. The edge is stored in the source question's language graphs. Its target can belong to another language graph.

For example, Q1 has tags matching both Python and Java, so its discussion is represented in both graphs. Q2 belongs to Java and is the target of a source link from Q1. The following statements illustrate this arrangement:

Q1 stands for http://stackoverflow.com/q/1 in both graphs; Q2 stands for http://stackoverflow.com/q/2. The two Q1 entries describe the same entity. There is no additional "same entity" edge. The two so:linkedTo statements are copies of the same source link stored in different named graphs.

These IDs are illustrative. Q2's reference in the Python graph does not by itself include Q2's title, body or answers there. Those descriptions are available when the package containing Q2's discussion has been loaded.

Each language package contains the selected language's discussion shards plus all shared tag-documentation shards and the schema. A Java graph appearing in a Python package's inventory can therefore contain shared documentation without containing Java's full question-and-answer collection. The same applies to extra-languages: its shared documentation can be present without its discussion shards. A referenced question can have no local description when its discussion is only in another package.

To follow these connections across languages, load the relevant packages from the same export identity into one RDF dataset and preserve the named graphs. Shared shards occur in multiple packages; avoid importing them repeatedly into stores that retain duplicate statements. For entity counts across graphs, use COUNT(DISTINCT ?question). Summing quad counts measures graph statements, including repeated statements in different language graphs, rather than unique discussions.

publication.json records each file's path, stored byte count, SHA-256 checksum and graph key, together with the export identity. RDF files are under graphs/. Internal run manifests, QLever indexes, credentials and uploader source are excluded from the archive.

Loading and querying

Load the .nq files as N-Quads and preserve their embedded named graph IRIs. For .nq.gz files, decompress them first if your RDF store cannot read gzip.

The following query counts loaded quads by graph. Use the graph inventory above to check the result, or restrict queries to http://stackoverflow.com/extra-languages when working with the selected language.

SELECT ?graph (COUNT(*) AS ?quads)
WHERE { GRAPH ?graph { ?s ?p ?o } }
GROUP BY ?graph

Retrieve questions and their available answers in this language graph:

PREFIX so: <http://stackoverflow.com/schema#>

SELECT ?question ?title ?answer
WHERE {
  GRAPH <http://stackoverflow.com/extra-languages> {
    ?question a so:Question .
    OPTIONAL { ?question so:hasTitle ?title }
    OPTIONAL { ?question so:hasAnswer ?answer }
  }
}
LIMIT 25

After loading other language packages, find linked or duplicate questions whose descriptions are available in another named graph:

PREFIX so: <http://stackoverflow.com/schema#>

SELECT DISTINCT ?question ?relation ?target ?targetGraph ?targetTitle
WHERE {
  GRAPH <http://stackoverflow.com/extra-languages> {
    VALUES ?relation { so:linkedTo so:duplicateOf }
    ?question a so:Question ; ?relation ?target .
  }
  GRAPH ?targetGraph {
    ?target a so:Question ; so:hasTitle ?targetTitle .
  }
  FILTER (?targetGraph != <http://stackoverflow.com/extra-languages>)
}
LIMIT 25

The second query returns only targets with loaded descriptions and titles. A target can occur in several graphs through overlapping tag selections; this does not mean the source and target questions use different languages exclusively.

Uses and limitations

The graph supports SPARQL analysis and retrieval over Stack Overflow discussions, including retrieval-augmented question answering. It represents the source snapshot; later edits and deletions on Stack Overflow are not reflected here.

Quality notes

• Votes are represented as aggregated upvote and downvote counts. Individual vote events and voter identities are not included.
• The graph includes selected moderation-related fields, such as a question's closure date, but does not represent the full moderation context or revision history.
• Community-generated Stack Overflow content can contain errors, outdated information and bias. Changes in languages, libraries and development practices can make advice less applicable over time.
• Scores, accepted-answer status and recent activity are source signals; they do not establish that an answer is correct or current.

License and attribution

This dataset is derived from Stack Overflow content. Its declared dataset license is CC BY-SA 4.0 (CC-BY-SA-4.0). Preserve required attribution and license notices when redistributing derived artifacts, and identify changes where required.

Individual source contributions retain their applicable license terms. Stack Overflow's licensing guidance distinguishes CC BY-SA 2.5, 3.0 and 4.0 according to contribution date. A dataset-level license declaration does not replace those source notices. The graph preserves source-provided license labels in so:contentLicense where available. Consult the original posts and revision records for author attribution and applicable terms. The publication's MDC datasheet provides its license declaration and contact details.

Citation

If you use this dataset, cite the Stack2Graph project. The research citation accompanies the required attribution for the underlying Stack Overflow content.

Kleybolte, L. A., Ventura, V., & Zarcone, A. (2026). Stack2Graph: A Structured Knowledge Representation of Stack Overflow Data for Retrieval-based Question Answering. In Proceedings of the Knowledge Graphs and Large Language Models Workshop (KG-LLM) @ LREC26 (pp. 120-133). European Language Resources Association (ELRA). https://doi.org/10.63317/583q4d3mwbdu.

Export provenance

These fingerprints identify the inputs and exporter used for this release. The export identity describes the build configuration and inputs; file checksums are listed separately in publication.json.

Source code and documentation: stack2graph.