Semantic network

• 1. Introduction

  ›

  Semantic networks map how concepts relate to one another, turning language and knowledge into a structure a machine can navigate. At their core they are graphs: vertices that stand for concepts, edges that carry the meaning of the relationship between them. That deceptively simple idea has been put to work in everything from machine translation to plagiarism detection, from social media analysis to search engines used by billions of people. How did a tool described in the third century AD by a Greek philosopher become a pillar of modern computing? And why did the boundary between a semantic network and something called a knowledge graph quietly dissolve over the course of a few decades?

• 2. Ancient Roots

  ›

  The Greek philosopher Porphyry wrote his commentary on Aristotle's categories in the third century AD, and in doing so produced what scholars now regard as the earliest documented use of a directed acyclic graph as a mnemonic tool. Porphyry was not building a database. He was trying to clarify how categories of being nest inside one another, which is exactly the kind of problem semantic networks were later designed to solve.

  Centuries of similar thinking in logic kept the underlying idea alive, even without a name for it. Thinkers working on how propositions connect, how meanings branch, and how one concept can contain or exclude another were all circling the same structure. The vocabulary would catch up much later, once computers gave the structure a practical home.

• 3. Computing Arrives

  ›

  Richard H. Richens of the Cambridge Language Research Unit was the first to implement what he called Semantic Nets for computers, in 1956. His goal was an interlingua: a representation of meaning that could sit between two natural languages and allow machine translation to proceed without depending on direct word-for-word substitution. The importance of that work, and of the Cambridge Language Research Unit itself, was only recognized belatedly by the wider field.

  Working independently, Robert F. Simmons and Sheldon Klein took a different path, grounding their implementation in first-order predicate calculus. Their starting point was a demonstration by Victor Yngve, the first President of the Association for Computational Linguistics, who had published descriptions of algorithms in 1960 for generating syntactically well-formed nonsense sentences using phrase structure grammar. Simmons later described how he and Klein, around 1962-1964, were fascinated by the technique and generalized it into a method that respected the semantic dependencies of words as they occurred in actual text. Other researchers, most notably M. Ross Quillian and others at System Development Corporation, contributed through the SYNTHEX project in the early 1960s. It is from publications at that organization that most modern uses of the term semantic network trace their lineage. Collins and Quillian, and later Collins and Loftus Quillian, produced prominent works that built on that foundation.

• 4. Knowledge Graphs Emerge

  ›

  In the late 1980s, two Dutch universities, Groningen and Twente, launched a joint project they called Knowledge Graphs. The core innovation was a constraint: edges in the graph had to be drawn from a limited, predefined set of possible relation types. That restriction made it possible to run algebraic operations on the graph, something that open-ended semantic networks could not easily support. Over the following decades the distinction between that constrained form and semantic networks more broadly became blurred rather than sharp.

  In 2012, Google applied the term Knowledge Graph to its own product, giving a technical lineage a brand name that became widely known outside academic circles. The Knowledge Graph proposed by Google in 2012 is, technically, an application of semantic network principles inside a search engine.

• 5. WordNet and Formal Languages

  ›

  WordNet is a lexical database of English that groups words into sets of synonyms called synsets, supplies short general definitions, and records the semantic relations between those sets. It gives precise names to relationships that everyday speakers use without thinking: meronymy, where A is part of B; holonymy, where B contains A; hyponymy, where A is a kind of B; hypernymy, where A is superordinate to B; synonymy; and antonymy. Researchers studying WordNet from a network-theory perspective found that it, along with semantic networks derived from Roget's Thesaurus and word association tasks, all share what network scientists call a small-world structure.

  Charles Sanders Peirce developed existential graphs, and John F. Sowa built on that tradition with conceptual graphs. Both approaches offer expressive power equal to or exceeding standard first-order predicate logic, which sets them apart from lexical networks like WordNet. Automated reasoners that use these representations can perform reliable logical deduction, exploiting the graph-theoretic features of the network during processing.

• 6. Specialized Forms

  ›

  Fawsy Bendeck's doctoral thesis, completed in 2008, formalized the Semantic Similarity Network, a specialized variant that contains dedicated relationships and propagation algorithms designed to simplify the calculation of semantic similarity. The same decade saw Herman Helbig complete a full description of MultiNet in 2006, a paradigm especially suited to the semantic representation of natural language expressions, and now used in several natural language processing applications alongside Stuart C. Shapiro's Semantic Network Processing System, known as SNePS.

  The Semantic Link Network followed a different trajectory. Its systematic theory and model were published in 2004, though the research direction traces to work on inheritance rules for efficient model retrieval from 1998 and to the Active Document Framework. Since 2003 the work developed toward social semantic networking, and it has since been extended from cyberspace into what researchers call cyber-physical-social space, where competition relations and symbiosis relations and their roles in an evolving society became active topics of study.

• 7. Practical Applications

  ›

  Natural language processing relies on semantic networks for tasks including semantic parsing and word-sense disambiguation, the problem of determining which meaning of an ambiguous word applies in a given sentence. Researchers also use them to analyze large text corpora and identify main themes, to surface biases in news coverage, and to map entire research fields at once.

  Embedding is a more recent technique that places multi-relational data like a semantic network into a low-dimensional mathematical space. That representation makes it possible to express entity relationships efficiently and to extract relations from text. Approaches include Bayesian clustering frameworks, energy-based frameworks, and TransE, a method that appeared at the NIPS 2013 conference. Applications of embedding knowledge-base data stretch into social network analysis and relationship extraction. Semantic networks are also applied in specialized information retrieval, such as plagiarism detection, where hierarchical relation data enables semantic compression that lets a system match word meanings independent of the specific words a writer chose.