Technique

A machine-readable pattern of lines that encodes a number, patented by Woodland and Silver in 1952 and put on groceries from 1974; it gave everyday objects identifiers that machines could read.

Text containing links that let a reader jump to other text. Ted Nelson coined the word in the 1960s; the idea goes back to the memex.

Including part of one document in another by reference, so the same content appears in both places and stays linked to its source. Ted Nelson coined the term in 1980.

Classification that emerges from many users freely tagging content; the term was coined by Thomas Vander Wal in 2004 and became the bottom-up counterpoint to formal ontologies.

Smallest unit of publishable scientific information: a single assertion written as RDF named graphs, bundled with its provenance and publication details so it can be cited and credited on its own.

Free tagging by users combined with volunteers who link synonymous tags to canonical ones, keeping the users' words while adding a controlled structure; named by Julia Bullard.

Retrieval-augmented generation in which a language model answers from a knowledge graph built over the source texts, rather than from isolated text chunks.

Naming data by a cryptographic hash of its content instead of by where it is stored. It is the main alternative to location-based URIs: the name says what a thing is, not where it lives.

Agreed list of terms for indexing and retrieval, so that one subject is always described the same way.

Machine-readable data inside ordinary HTML pages, written as microformats, RDFa, microdata or JSON-LD and usually using the schema.org vocabulary.

Classification that describes a subject by combining independent facets, such as topic, place and time, instead of placing it in one fixed hierarchy.

Plain JSON over HTTP with each API defining its own fields, without shared vocabularies or global identifiers; the everyday alternative most developers chose over Linked Data.

Neural language model trained on very large text collections; since the early 2020s the main way software works with unstructured text.

Data that describes other data or documents, such as author, date, subject or format.

AI that combines neural networks with symbolic knowledge such as rules and knowledge graphs; it drew new attention in the 2020s as a way to ground large language models.

Storage that a person controls, from which applications read and write data with permission, instead of each application keeping its own copy.

Technique in which a language model first retrieves relevant documents or data and then generates its answer from them.

Classification of things or concepts into a hierarchy of categories.

Controlled vocabulary whose terms are linked by broader, narrower, related and preferred-term relations.

Database built to store RDF triples and answer SPARQL queries; Virtuoso is one example.

Ted Nelson's paper at the 1965 ACM national conference, which introduced the word 'hypertext' in print.

Brin and Page's 1998 paper describing Google and PageRank, which ranks pages by the links pointing to them; it showed that the Web's link structure could serve as a signal of authority, with no metadata needed from publishers.

Step-by-step guide by Natasha Noy and Deborah McGuinness to building an ontology, with examples in Protégé.

WWW 2005 paper extending RDF so that a set of statements can be named by a URI and described like any resource. It shows publishers signing their graphs and consumers deciding which graphs to trust.

Roy Fielding's June 2005 message announcing the W3C TAG's answer to httpRange-14: a URI that answers with a 2xx response names an information resource, while one that answers with a 303 redirect may name anything.

Book by Jérôme Euzenat and Pavel Shvaiko (2007; second edition 2013) on methods for finding correspondences between the terms of different ontologies, so that data described in one can be used with another.

W3C Interest Group Note (2008) giving publishers two recipes for URIs that name things rather than pages: hash URIs, and 303 redirects based on the httpRange-14 resolution.

Practitioner's guide by Dean Allemang and Jim Hendler to modelling with RDFS and OWL.

2009 ESWC paper by Georgi Kobilarov, Tom Scott, Yves Raimond and colleagues at the BBC and Freie Universität Berlin, describing how BBC Programmes and BBC Music used DBpedia and MusicBrainz identifiers to connect content across the BBC's site.

2009 workshop paper by Oktie Hassanzadeh and Mariano Consens presenting LinkedMDB, linked open data about films and their links to other datasets.

ISWC 2009 paper by Hartig, Bizer and Freytag proposing link-traversal querying: a SPARQL query is answered by following links and looking up URIs as it runs, so the data sources need not be known in advance.

2010 paper by Paul Groth, Andrew Gibson and Jan Velterop in Information Services & Use defining the nanopublication.

2012 NIPS paper by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton whose deep network, trained on GPUs, won the ImageNet image-recognition contest by a wide margin and turned machine learning toward deep learning.

2013 NIPS paper introducing TransE, which represents each entity in a knowledge graph as a vector and each relation as a translation between vectors, so that missing facts can be predicted; the starting point of knowledge-graph embeddings.

2013 paper by Tomas Mikolov, Kai Chen, Greg Corrado and Jeffrey Dean introducing word2vec: vectors learned from which words occur together, in which relations such as king to queen show up as directions.

SEMANTiCS 2014 paper introducing Shape Expressions (ShEx), a structural schema language that describes the shape RDF data should have and checks data against it.

2016 Journal of Web Semantics paper proposing Triple Pattern Fragments: servers answer only simple triple patterns and clients do the rest of the query, trading some speed for servers that are cheap to run and stay up.

Semantic Web journal paper describing LOV, a curated catalogue of reusable vocabularies that records how they link to each other, their versions and who maintains them, to help publishers choose existing vocabularies instead of inventing new ones.

2017 paper by Ashish Vaswani and seven Google colleagues introducing the transformer, a network built on attention that became the basis of large language models.

2018 paper by Julia Bullard in Knowledge Organization comparing three user-generated collections that structure their tags through human judgment, and proposing the curated folksonomy as an object of study.

2018 ESWC paper that collected about 559 million owl:sameAs statements from a 2015 crawl of Linked Data, computed which names they make equivalent, and published the result with an error score for each link.

2018 paper by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova: a transformer pre-trained on large amounts of text that set new results on many language tasks; Google applied it to Search in 2019.

Microsoft Research paper by Darren Edge and nine co-authors introducing the GraphRAG method.

2024 paper by Liu, Semnani, Triedman, Xu, Zhao and Lam with a dataset of real questions from Wikidata's 'Request a Query' forum, where users ask others to write SPARQL for them, and a language-model agent that writes such queries the way an expert would.